Add 9 LoRA(s): detailer-ce, instagram-reality, jacques-tati-movies-style, cg-texture-master-iray-skin-rendering-cg-3s…

.gitattributes

@@ -1413,3 +1413,58 @@ ignacia/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text
 ignacia/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text
 ignacia/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text
 ignacia/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text
+detailer-ce/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text
+detailer-ce/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text
+detailer-ce/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text
+detailer-ce/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text
+detailer-ce/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text
+detailer-ce/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text
+instagram-reality/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text
+instagram-reality/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text
+instagram-reality/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text
+instagram-reality/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text
+instagram-reality/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text
+instagram-reality/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text
+jacques-tati-movies-style/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text
+jacques-tati-movies-style/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text
+jacques-tati-movies-style/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text
+jacques-tati-movies-style/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text
+jacques-tati-movies-style/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text
+jacques-tati-movies-style/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text
+cg-texture-master-iray-skin-rendering-cg-3s/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text
+cg-texture-master-iray-skin-rendering-cg-3s/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text
+cg-texture-master-iray-skin-rendering-cg-3s/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text
+cg-texture-master-iray-skin-rendering-cg-3s/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text
+cg-texture-master-iray-skin-rendering-cg-3s/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text
+cg-texture-master-iray-skin-rendering-cg-3s/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text
+klein-realistic-comic-slider/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text
+klein-realistic-comic-slider/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text
+klein-realistic-comic-slider/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text
+klein-realistic-comic-slider/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text
+klein-realistic-comic-slider/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text
+klein-realistic-comic-slider/example_images/06_poster.jpg filter=lfs diff=lfs merge=lfs -text
+klein-realistic-comic-slider/example_images/06_video.mp4 filter=lfs diff=lfs merge=lfs -text
+claire-redfield-resident-evil-2-remake-in-real-life/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text
+claire-redfield-resident-evil-2-remake-in-real-life/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text
+claire-redfield-resident-evil-2-remake-in-real-life/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text
+claire-redfield-resident-evil-2-remake-in-real-life/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text
+claire-redfield-resident-evil-2-remake-in-real-life/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text
+claire-redfield-resident-evil-2-remake-in-real-life/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text
+aniedit-flux-2-klein/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text
+aniedit-flux-2-klein/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text
+aniedit-flux-2-klein/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text
+aniedit-flux-2-klein/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text
+aniedit-flux-2-klein/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text
+aniedit-flux-2-klein/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text
+hand-drawn-sketch/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text
+hand-drawn-sketch/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text
+hand-drawn-sketch/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text
+hand-drawn-sketch/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text
+hand-drawn-sketch/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text
+hand-drawn-sketch/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text
+the-body-sfw/example_images/01_image.jpeg filter=lfs diff=lfs merge=lfs -text
+the-body-sfw/example_images/02_image.jpeg filter=lfs diff=lfs merge=lfs -text
+the-body-sfw/example_images/03_image.jpeg filter=lfs diff=lfs merge=lfs -text
+the-body-sfw/example_images/04_image.jpeg filter=lfs diff=lfs merge=lfs -text
+the-body-sfw/example_images/05_image.jpeg filter=lfs diff=lfs merge=lfs -text
+the-body-sfw/example_images/06_image.jpeg filter=lfs diff=lfs merge=lfs -text

aniedit-flux-2-klein/AniEdit9B_v2.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d37d58268f6e704f7d5d5daf3cfc4d69bc82189c71e17494b5e02b6892274927
+size 994079944

aniedit-flux-2-klein/info.txt

@@ -0,0 +1,247 @@
+═══════════════════════════════════════════════════════════════════════════════
+  ANIEDIT (FLUX 2 KLEIN)
+═══════════════════════════════════════════════════════════════════════════════
+
+───────────────────────────────────────────────────────────────────────────────
+  IDENTITY
+───────────────────────────────────────────────────────────────────────────────
+  Base model        FLUX.2 [klein] 9B
+  Generation mode   not declared by the author
+  Checkpoint        not determined — verify before use
+  Category          Visual style
+  Author            memnoid
+  Source            https://civitai.com/models/2332320
+  Provenance        CivitAI · model 2332320 · version 2642969
+
+───────────────────────────────────────────────────────────────────────────────
+  WEIGHTS
+───────────────────────────────────────────────────────────────────────────────
+  File              AniEdit9B_v2.safetensors
+  Version           9B v2
+  Size              948.0 MB
+  Format            SafeTensor / unknown precision
+  SHA256            d37d58268f6e704f7d5d5daf3cfc4d69bc82189c71e17494b5e02b6892274927
+  Published         2026-01-30
+
+───────────────────────────────────────────────────────────────────────────────
+  TRIGGER WORDS
+───────────────────────────────────────────────────────────────────────────────
+  No trigger word. The LoRA applies as soon as it is loaded —
+  strength is the only way to dial the effect up or down.
+
+───────────────────────────────────────────────────────────────────────────────
+  RECOMMENDED SETTINGS
+───────────────────────────────────────────────────────────────────────────────
+  The author did not publish specific settings.
+  No author settings were published. Community baseline: steps Distilled
+  FLUX.2-klein-9B: 4 (HF card example; it is step-distilled to exactly this,
+  and more steps do not refine further). Base FLUX.2-klein-base-9B: 50 in the
+  HF card example and in ComfyUI's benchmark; community practice runs lower,
+  with ~20 for production and 8-12 for fast iteration reported in the CivitAI
+  consistency-LoRA article., cfg Distilled: guidance_scale 1.0 (HF card). It
+  is guidance-distilled and its transformer config sets
+  guidance_embeds=false, so raising CFG degrades rather than improves. Base:
+  guidance_scale 4.0 (HF card); the CivitAI consistency-LoRA article
+  recommends 3.5-4.5 for most edits., sampler Not stated by BFL or by
+  docs.comfy.org, and the official flux2 GitHub repo documents no sampler,
+  scheduler or shift values either. Community consensus is euler with a
+  resolution-dependent shifted-Euler / flow-match schedule (ComfyUI's Flux2
+  workflows use a scheduler that interpolates shift between base_shift and
+  max_shift by output resolution). Start from ComfyUI's downloadable Flux.2
+  klein workflow template rather than a blog. See unverified., resolution
+  1024x1024 in every BFL HF card example and in CivitAI's 9B training recipe
+  ("resolution": 1024). Non-square ratios work but no official list of
+  supported dimensions or a stated alignment multiple was found — see
+  unverified., lora_strength 1.0 is the diffusers/ComfyUI default and what
+  CivitAI applies to training previews (strength 1.0). Published per-LoRA
+  guidance varies and should be read from the LoRA, not assumed: the widely
+  used Klein 9B consistency LoRA's CivitAI article says start at 0.5 within a
+  0.2-1.0 range (0.2-0.4 loose, 0.5-0.7 balanced, 0.8-1.0 maximum structural
+  stability) and warns that too high a strength stops edits from applying;
+  SOLRICKS' Realistic Detail LoRA card says 0.6-0.8 subtle, 0.8-1.0 stronger,
+  start 0.8; the vafipas663 distillation-delta LoRA card warns it "produces
+  static noise at strength > 0.5. And sometimes, you have to go as low as
+  0.1.".
+
+───────────────────────────────────────────────────────────────────────────────
+  ADOPTION & POPULARITY
+───────────────────────────────────────────────────────────────────────────────
+  Downloads               2,840
+  Rating                  100% positive (128/128)
+  Comments                15
+  Published               218 days ago
+  Download rate           8.9/day since release
+
+  Adoption rank           #103 of 200 in this collection  (Modest)
+  Momentum rank           #68 of 200 by download rate  (Steady)
+
+  Adoption tier is the percentile of total downloads within this collection;
+  momentum is the percentile of downloads-per-day since release. Momentum is
+  ranked rather than measured against a fixed rate, because this base model
+  is itself new and every LoRA looks fast on an absolute scale. Both numbers
+  are CivitAI's, read on the scrape date at the foot of this file.
+
+───────────────────────────────────────────────────────────────────────────────
+  CHAINING / STACKING
+───────────────────────────────────────────────────────────────────────────────
+  Stacks with others      not stated
+
+  The author gave no chaining guidance. The rules below are the
+  general FLUX.2 [klein] 9B ones, not anything specific to this LoRA.
+
+  General FLUX.2 [klein] 9B rules that apply here (see REFERENCE.txt):
+
+  • Chain LoRAs by stringing LoraLoaderModelOnly nodes in series, model
+    output into model input. There is no CLIP side to route — klein LoRAs
+    are model-only because the Qwen3 text encoder is not trained (CivitAI's
+    recipe fixes trainTextEncoder: false).
+  • Only stack LoRAs of the same parameter size. 9B stacks with 9B; a 4B
+    LoRA in a 9B stack is an architectural mismatch (3072 vs 4096 hidden,
+    5+20 vs 8+24 blocks), not a quality question.
+  • Mixing a "Flux.2 Klein 9B" LoRA and a "Flux.2 Klein 9B-base" LoRA in
+    one stack is fine — those strings describe the training checkpoint, not
+    incompatible tensor layouts. Set the sampler for the checkpoint
+    actually loaded.
+  • Run base-trained LoRAs on the distilled checkpoint when you want speed:
+    HF's klein LoRA guide reports the adapter loads there and typically
+    gives better results than on base, at 4 steps.
+  • A base-to-turbo / distillation-extraction LoRA is a whole-model delta,
+    not a content LoRA. Put it on the base checkpoint only, and treat its
+    strength as a steps/CFG dial rather than a style knob. Do not add it on
+    top of the already-distilled checkpoint.
+  • Read each LoRA's own published strength range instead of defaulting
+    everything to 1.0 — klein LoRAs in the wild ship very different ones
+    (0.5 start for the consistency LoRA, 0.8 for SOLRICKS' detail LoRA,
+    under 0.5 and sometimes 0.1 for the vafipas663 distillation delta).
+  • With a consistency or edit LoRA in the stack, excessive strength
+    suppresses the prompt: the consistency LoRA's article says that if
+    edits stop applying, lower it before changing anything else.
+  • Prefer runtime LoRA patching over baking a LoRA into the checkpoint:
+    ComfyUI's Model Merge / Model Save nodes have not detected Flux2 Klein
+    models (4B or 9B) since a March 2026 update — issue #13637, still open.
+  • The author never states which weight family this targets. Two different
+    questions, two different answers. Within 9B (distilled vs base — the
+    two CivitAI strings): not a real incompatibility. Both load through the
+    same Flux2KleinPipeline and have the same repo layout; for the 4B pair,
+    where both transformer configs are public, the distil
+
+───────────────────────────────────────────────────────────────────────────────
+  PROS
+───────────────────────────────────────────────────────────────────────────────
+  + Well established — 2,840 downloads. [source: CivitAI stats]
+  + Positively received — 128 thumbs-up with no down-votes. [source:
+    CivitAI stats]
+
+───────────────────────────────────────────────────────────────────────────────
+  CONS & CAVEATS
+───────────────────────────────────────────────────────────────────────────────
+  - Author references NSFW output — review before using in shared work.
+    [source: description]
+  - No trigger word — the LoRA is always on once loaded, so strength is the
+    only control. [source: model metadata]
+  - Author never states which generation mode it was trained for; test
+    against your own workflow before relying on it. [source: description]
+  - Large at 948 MB — slower to load and heavier to stack. [source: file
+    size]
+
+───────────────────────────────────────────────────────────────────────────────
+  COMPATIBILITY WARNING
+───────────────────────────────────────────────────────────────────────────────
+  Two different questions, two different answers. Within 9B (distilled vs
+  base — the two CivitAI strings): not a real incompatibility. Both load
+  through the same Flux2KleinPipeline and have the same repo layout; for the
+  4B pair, where both transformer configs are public, the distilled and base
+  configs are byte-for-byte identical in every shape field, and the 9B pair
+  is built the same way. HF's own klein LoRA guide states the workflow
+  directly: "Distilled models are step-compressed for fast inference; you
+  train against the base checkpoint and the adapter still loads on the
+  distilled model afterward", and "Applying the LoRA on the distilled model
+  typically gives better results than the base model, and it's faster" (that
+  post is written against 4B, but the family structure is the same). So the
+  two CivitAI strings record which checkpoint the author trained or previewed
+  against, not a hard compatibility wall. The real risk is a silent settings
+  mismatch: load a base-trained LoRA on the distilled checkpoint while
+  keeping the author's 20-50 step / CFG ~4 preview settings and the output
+  degrades badly (distilled wants 4 steps / guidance 1.0), or the reverse.
+  Nothing errors — it just looks wrong. Across sizes (9B vs 4B): a genuine
+  architectural mismatch, from the published transformer configs — different
+  hidden width (4096 vs 3072), different block counts (8+24 vs 5+20) and a
+  different text-conditioning dim (12288 vs 7680, following Qwen3-8B vs
+  Qwen3-4B). A 4B LoRA cannot patch a 9B model: expect missing keys and shape
+  errors, not degraded output. The exact ComfyUI error text is unverified. KV
+  variant: same 9B size and same 4-step distillation, but a distinct
+  checkpoint with its own pipeline; whether standard klein 9B LoRAs patch it
+  cleanly is not documented — see unverified.
+
+───────────────────────────────────────────────────────────────────────────────
+  AUTHOR'S DESCRIPTION
+───────────────────────────────────────────────────────────────────────────────
+  A comprehensive finetune of Flux.2-Klein 4B/9B, that works on distilled
+  versions. Art style transfer using 4 reference images: Workflow
+  Upscale any anime image, prompt at 2-4MP: Workflow
+  ControlNet any single subject image: Workflow ...And many other edit tasks:
+  Real life to anime, change clothes, prompt new pose, change expression,
+  hairtstyle, pupils, Compose two subjects, convert to 3d model,
+  anthropomorphize, etc. Dataset (1500 image pairs CONTAINS NSFW /Drawn
+  nudity)
+
+───────────────────────────────────────────────────────────────────────────────
+  VERSION NOTES
+───────────────────────────────────────────────────────────────────────────────
+  Doubled the dataset.
+
+────────────────────────────────────────────────────���──────────────────────────
+  EXAMPLE MEDIA
+───────────────────────────────────────────────────────────────────────────────
+  6 preview(s) in ./example_images/ — community results for
+  this LoRA. Videos keep a matching _poster.jpg still frame.
+
+  No prompts are recorded below: CivitAI's public API flags these posts as
+  having generation metadata but does not return it, so the prompts cannot be
+  scraped. Open the model page to read them.
+
+  01_image.jpeg  (image, 832x1216)
+  02_image.jpeg  (image, 1056x1952)
+  03_image.jpeg  (image, 832x1216)
+  04_image.jpeg  (image, 1216x832)
+  05_image.jpeg  (image, 1248x1664)
+  06_image.jpeg  (image, 832x1216)
+
+───────────────────────────────────────────────────────────────────────────────
+  HOW TO USE (COMFYUI)
+───────────────────────────────────────────────────────────────────────────────
+  1. Diffusion model -> ComfyUI/models/diffusion_models/. Distilled:
+  flux-2-klein-9b-fp8.safetensors. Base: flux-2-klein-
+  base-9b-fp8.safetensors. (BF16 single-file weights
+  flux-2-klein-9b.safetensors / flux-2-klein-base-9b.safetensors from the
+  gated HF repos also work if you have ~29GB; community GGUF builds exist
+  from unsloth and leejet for lower-VRAM setups.)
+  2. Text encoder ->
+  ComfyUI/models/text_encoders/qwen_3_8b_fp8mixed.safetensors (from Comfy-
+  Org/vae-text-encorder-for-flux-klein-9b). This is the 9B family's encoder;
+  the 4B family uses qwen_3_4b.safetensors — do not mix them.
+  3. VAE -> ComfyUI/models/vae/flux2-vae.safetensors (shared across klein
+  sizes).
+  4. LoRA -> ComfyUI/models/loras/ as a plain .safetensors.
+  5. Load with Load Diffusion Model (UNETLoader) for the transformer,
+  CLIPLoader for the Qwen3 text encoder, VAELoader for the VAE.
+  docs.comfy.org ships prebuilt Flux.2 klein graphs under Workflow Templates
+  for both the distilled and base branches — start from those rather than
+  wiring by hand, since neither the docs nor BFL's repo state
+  sampler/scheduler/shift in prose; those values live in the template.
+  6. Insert LoraLoaderModelOnly between the diffusion-model loader and the
+  sampler. Model-only is correct for klein: the Qwen3 text encoder is not
+  trained by the standard klein trainers, so there is no text-encoder half to
+  patch.
+  7. Set steps and guidance to match the checkpoint you actually loaded, not
+  the one the LoRA was trained on: 4 steps / guidance 1.0 on distilled,
+  ~20-50 steps / guidance ~4.0 on base.
+  8. Note the 9B repos are gated on HuggingFace — you must accept the FLUX
+  Non-Commercial License on the repo page before any download, including
+  scripted ones.
+
+───────────────────────────────────────────────────────────────────────────────
+  Scraped by lorakit on 2026-09-05 03:55 UTC.
+  Stats and description are the author's; pros/cons are derived from the
+  evidence tagged beside each line.
+───────────────────────────────────────────────────────────────────────────────

aniedit-flux-2-klein/metadata.json

@@ -0,0 +1,131 @@
+{
+  "name": "AniEdit (Flux 2 Klein)",
+  "slug": "aniedit-flux-2-klein",
+  "base_model": "FLUX.2 [klein] 9B",
+  "media_type": "image",
+  "model_id": 2332320,
+  "version_id": 2642969,
+  "version_name": "9B v2",
+  "source": "civitai",
+  "url": "https://civitai.com/models/2332320",
+  "repo_id": null,
+  "file_path": null,
+  "author": "memnoid",
+  "category": {
+    "key": "style",
+    "label": "Visual style"
+  },
+  "modalities": [],
+  "primary_modality": "unspecified",
+  "checkpoint_family": "unknown",
+  "trigger_words": [],
+  "recommended_settings": {},
+  "chaining": {
+    "author": {
+      "stacks": "not stated",
+      "notes": "",
+      "avoid": [],
+      "requires": [],
+      "order": "",
+      "strength": "",
+      "stated": false
+    },
+    "general_rules": [
+      "Chain LoRAs by stringing LoraLoaderModelOnly nodes in series, model output into model input. There is no CLIP side to route — klein LoRAs are model-only because the Qwen3 text encoder is not trained (CivitAI's recipe fixes trainTextEncoder: false).",
+      "Only stack LoRAs of the same parameter size. 9B stacks with 9B; a 4B LoRA in a 9B stack is an architectural mismatch (3072 vs 4096 hidden, 5+20 vs 8+24 blocks), not a quality question.",
+      "Mixing a \"Flux.2 Klein 9B\" LoRA and a \"Flux.2 Klein 9B-base\" LoRA in one stack is fine — those strings describe the training checkpoint, not incompatible tensor layouts. Set the sampler for the checkpoint actually loaded.",
+      "Run base-trained LoRAs on the distilled checkpoint when you want speed: HF's klein LoRA guide reports the adapter loads there and typically gives better results than on base, at 4 steps.",
+      "A base-to-turbo / distillation-extraction LoRA is a whole-model delta, not a content LoRA. Put it on the base checkpoint only, and treat its strength as a steps/CFG dial rather than a style knob. Do not add it on top of the already-distilled checkpoint.",
+      "Read each LoRA's own published strength range instead of defaulting everything to 1.0 — klein LoRAs in the wild ship very different ones (0.5 start for the consistency LoRA, 0.8 for SOLRICKS' detail LoRA, under 0.5 and sometimes 0.1 for the vafipas663 distillation delta).",
+      "With a consistency or edit LoRA in the stack, excessive strength suppresses the prompt: the consistency LoRA's article says that if edits stop applying, lower it before changing anything else.",
+      "Prefer runtime LoRA patching over baking a LoRA into the checkpoint: ComfyUI's Model Merge / Model Save nodes have not detected Flux2 Klein models (4B or 9B) since a March 2026 update — issue #13637, still open.",
+      "The author never states which weight family this targets. Two different questions, two different answers. Within 9B (distilled vs base — the two CivitAI strings): not a real incompatibility. Both load through the same Flux2KleinPipeline and have the same repo layout; for the 4B pair, where both transformer configs are public, the distil"
+    ]
+  },
+  "pros": [
+    "Well established — 2,840 downloads. [source: CivitAI stats]",
+    "Positively received — 128 thumbs-up with no down-votes. [source: CivitAI stats]"
+  ],
+  "cons": [
+    "Author references NSFW output — review before using in shared work. [source: description]",
+    "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]",
+    "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]",
+    "Large at 948 MB — slower to load and heavier to stack. [source: file size]"
+  ],
+  "file": {
+    "name": "AniEdit9B_v2.safetensors",
+    "size_bytes": 994079944,
+    "size_mb": 948.03,
+    "sha256": "d37d58268f6e704f7d5d5daf3cfc4d69bc82189c71e17494b5e02b6892274927"
+  },
+  "stats": {
+    "downloads": 2840,
+    "thumbs_up": 128,
+    "thumbs_down": 0
+  },
+  "popularity": {
+    "downloads": 2840,
+    "thumbs_up": 128,
+    "thumbs_down": 0,
+    "comments": 15,
+    "approval": "100% positive (128/128)",
+    "age_days": 217.8,
+    "downloads_per_day": 8.89,
+    "rank_by_downloads": 103,
+    "rank_by_momentum": 68,
+    "of_total": 200,
+    "adoption_tier": "Modest",
+    "momentum": "Steady"
+  },
+  "selection_score": 71.3,
+  "selection_reasons": [
+    "2,840 downloads",
+    "128 thumbs-up",
+    "top-ranked for its category"
+  ],
+  "example_media": [
+    {
+      "file": "01_image.jpeg",
+      "type": "image",
+      "resolution": "832x1216",
+      "prompt": ""
+    },
+    {
+      "file": "02_image.jpeg",
+      "type": "image",
+      "resolution": "1056x1952",
+      "prompt": ""
+    },
+    {
+      "file": "03_image.jpeg",
+      "type": "image",
+      "resolution": "832x1216",
+      "prompt": ""
+    },
+    {
+      "file": "04_image.jpeg",
+      "type": "image",
+      "resolution": "1216x832",
+      "prompt": ""
+    },
+    {
+      "file": "05_image.jpeg",
+      "type": "image",
+      "resolution": "1248x1664",
+      "prompt": ""
+    },
+    {
+      "file": "06_image.jpeg",
+      "type": "image",
+      "resolution": "832x1216",
+      "prompt": ""
+    }
+  ],
+  "tags": [
+    "base model",
+    "edit",
+    "anime"
+  ],
+  "scraped_at": "2026-09-05T03:55:52+00:00",
+  "path": "flux2-klein-9b/aniedit-flux-2-klein"
+}

cg-texture-master-iray-skin-rendering-cg-3s/info.txt

@@ -0,0 +1,225 @@
+═══════════════════════════════════════════════════════════════════════════════
+  CG TEXTURE MASTER & IRAY SKIN RENDERING | 极致CG质感与3S皮肤材质
+═══════════════════════════════════════════════════════════════════════════════
+
+───────────────────────────────────────────────────────────────────────────────
+  IDENTITY
+───────────────────────────────────────────────────────────────────────────────
+  Base model        FLUX.2 [klein] 9B
+  Generation mode   Text-to-image
+  Checkpoint        FLUX.2-klein-9B (distilled) — CivitAI string "Flux.2 Klein 9B"
+  Category          Visual style
+  Author            nnegret
+  Source            https://civitai.com/models/2388820
+  Provenance        CivitAI · model 2388820 · version 2686105
+
+───────────────────────────────────────────────────────────────────────────────
+  WEIGHTS
+───────────────────────────────────────────────────────────────────────────────
+  File              nnegrat_cgstyle_9b_1800.safetensors
+  Version           Klein9B_V1
+  Size              158.0 MB
+  Format            SafeTensor / unknown precision
+  SHA256            d58f7c9ffb43e3c3cd9f72adbec905f06397218fd885aa86e4e90e96def8b980
+  Published         2026-02-13
+
+───────────────────────────────────────────────────────────────────────────────
+  TRIGGER WORDS
+───────────────────────────────────────────────────────────────────────────────
+  › "artwork in cgstyle_9b"
+
+───────────────────────────────────────────────────────────────────────────────
+  RECOMMENDED SETTINGS
+───────────────────────────────────────────────────────────────────────────────
+  Taken verbatim from the author's description:
+
+  Sampler             euler
+
+───────────────────────────────────────────────────────────────────────────────
+  ADOPTION & POPULARITY
+───────────────────────────────────────────────────────────────────────────────
+  Downloads               1,579
+  Rating                  100% positive (136/136)
+  Comments                0
+  Published               204 days ago
+  Download rate           4.0/day since release
+
+  Adoption rank           #141 of 200 in this collection  (Modest)
+  Momentum rank           #141 of 200 by download rate  (Long tail)
+
+  Adoption tier is the percentile of total downloads within this collection;
+  momentum is the percentile of downloads-per-day since release. Momentum is
+  ranked rather than measured against a fixed rate, because this base model
+  is itself new and every LoRA looks fast on an absolute scale. Both numbers
+  are CivitAI's, read on the scrape date at the foot of this file.
+
+───────────────────────────────────────────────────────────────────────────────
+  CHAINING / STACKING
+───────────────────────────────────────────────────────────────────────────────
+  Stacks with others      not stated
+
+  The author gave no chaining guidance. The rules below are the
+  general FLUX.2 [klein] 9B ones, not anything specific to this LoRA.
+
+  General FLUX.2 [klein] 9B rules that apply here (see REFERENCE.txt):
+
+  • Chain LoRAs by stringing LoraLoaderModelOnly nodes in series, model
+    output into model input. There is no CLIP side to route — klein LoRAs
+    are model-only because the Qwen3 text encoder is not trained (CivitAI's
+    recipe fixes trainTextEncoder: false).
+  • Only stack LoRAs of the same parameter size. 9B stacks with 9B; a 4B
+    LoRA in a 9B stack is an architectural mismatch (3072 vs 4096 hidden,
+    5+20 vs 8+24 blocks), not a quality question.
+  • Mixing a "Flux.2 Klein 9B" LoRA and a "Flux.2 Klein 9B-base" LoRA in
+    one stack is fine — those strings describe the training checkpoint, not
+    incompatible tensor layouts. Set the sampler for the checkpoint
+    actually loaded.
+  • Run base-trained LoRAs on the distilled checkpoint when you want speed:
+    HF's klein LoRA guide reports the adapter loads there and typically
+    gives better results than on base, at 4 steps.
+  • A base-to-turbo / distillation-extraction LoRA is a whole-model delta,
+    not a content LoRA. Put it on the base checkpoint only, and treat its
+    strength as a steps/CFG dial rather than a style knob. Do not add it on
+    top of the already-distilled checkpoint.
+  • Read each LoRA's own published strength range instead of defaulting
+    everything to 1.0 — klein LoRAs in the wild ship very different ones
+    (0.5 start for the consistency LoRA, 0.8 for SOLRICKS' detail LoRA,
+    under 0.5 and sometimes 0.1 for the vafipas663 distillation delta).
+  • With a consistency or edit LoRA in the stack, excessive strength
+    suppresses the prompt: the consistency LoRA's article says that if
+    edits stop applying, lower it before changing anything else.
+  • Prefer runtime LoRA patching over baking a LoRA into the checkpoint:
+    ComfyUI's Model Merge / Model Save nodes have not detected Flux2 Klein
+    models (4B or 9B) since a March 2026 update — issue #13637, still open.
+  • Step-distilled to exactly 4 inference steps and guidance-distilled; HF
+    card example is num_inference_steps=4, guidance_scale=1.0. Its
+    transformer config has guidance_embeds=false. This is the intended
+    inference target: ComfyUI's blog benchmarks it at ~2s and 19.6GB VRAM
+    on an RTX 5090. Not offered as a training target by BFL's docs or
+    CivitAI's recipe.
+
+───────────────────────────────────────────────────────────────────────────────
+  PROS
+───────────────────────────────────────────────────────────────────────────────
+  + Well established — 1,579 downloads. [source: CivitAI stats]
+  + Positively received — 136 thumbs-up with no down-votes. [source:
+    CivitAI stats]
+  + Explicit trigger word(s) — artwork in cgstyle_9b — so the effect can be
+    turned on and off from the prompt. [source: model metadata]
+  + Compact at 158 MB — cheap to stack with other LoRAs. [source: file
+    size]
+  + Author published concrete recommended settings. [source: description]
+
+───────────────────────────────────────────────────────────────────────────────
+  CONS & CAVEATS
+───────────────────────────────────────────────────────────────────────────────
+  (none recorded)
+
+───────────────────────────────────────────────────────────────────────────────
+  COMPATIBILITY WARNING
+───────────────────────────────────────────────────────────────────────────────
+  Two different questions, two different answers. Within 9B (distilled vs
+  base — the two CivitAI strings): not a real incompatibility. Both load
+  through the same Flux2KleinPipeline and have the same repo layout; for the
+  4B pair, where both transformer configs are public, the distilled and base
+  configs are byte-for-byte identical in every shape field, and the 9B pair
+  is built the same way. HF's own klein LoRA guide states the workflow
+  directly: "Distilled models are step-compressed for fast inference; you
+  train against the base checkpoint and the adapter still loads on the
+  distilled model afterward", and "Applying the LoRA on the distilled model
+  typically gives better results than the base model, and it's faster" (that
+  post is written against 4B, but the family structure is the same). So the
+  two CivitAI strings record which checkpoint the author trained or previewed
+  against, not a hard compatibility wall. The real risk is a silent settings
+  mismatch: load a base-trained LoRA on the distilled checkpoint while
+  keeping the author's 20-50 step / CFG ~4 preview settings and the output
+  degrades badly (distilled wants 4 steps / guidance 1.0), or the reverse.
+  Nothing errors — it just looks wrong. Across sizes (9B vs 4B): a genuine
+  architectural mismatch, from the published transformer configs — different
+  hidden width (4096 vs 3072), different block counts (8+24 vs 5+20) and a
+  different text-conditioning dim (12288 vs 7680, following Qwen3-8B vs
+  Qwen3-4B). A 4B LoRA cannot patch a 9B model: expect missing keys and shape
+  errors, not degraded output. The exact ComfyUI error text is unverified. KV
+  variant: same 9B size and same 4-step distillation, but a distinct
+  checkpoint with its own pipeline; whether standard klein 9B LoRAs patch it
+  cleanly is not documented — see unverified.
+
+───────────────────────────────────────────────────────────────────────────────
+  AUTHOR'S DESCRIPTION
+───────────────────────────────────────────────────────────────────────────────
+  交流Q群：173273201 核心特点 / Key Features - Iray Style Rendering:
+  带来真实的物理光影感，画面通透且富有深度。 / Provides a realistic physical sense of light and
+  shadow, making the image transparent and deep.
+
+  - Ultimate Skin Texture: 重点优化了皮肤的 3S 材质，表现出通透、红润且极其光滑的皮肤触感。 / Specifically
+  optimized for SSS materials, showcasing translucent, rosy, and extremely
+  smooth skin.
+
+  - Versatile Transformation: 无论是动漫转 CG 还是真人转 CG，都能保持极高的风格一致性。 / Whether it's
+  Anime-to-CG or Realistic-to-CG, it maintains high style consistency.
+
+  - Multi-Scenario Support: 完美支持文生图 (T2I) 和图生图 (I2I)。 / Fully compatible with
+  both Text-to-Image and Image-to-Image workflows.
+
+  使用建议 / Usage Recommendations - 权重 / LoRA Weight: * Range: 0.75 - 0.9
+  - Recommended: 0.85
+
+  - 采样器 / Sampler: er_sde （stabilize） euler (dynamic)
+
+  示例展示 / Gallery Tips
+
+───────────────────────────────────────────────────────────────────────────────
+  EXAMPLE MEDIA
+───────────────────────────────────────────────────────────────────────────────
+  6 preview(s) in ./example_images/ — community results for
+  this LoRA. Videos keep a matching _poster.jpg still frame.
+
+  No prompts are recorded below: CivitAI's public API flags these posts as
+  having generation metadata but does not return it, so the prompts cannot be
+  scraped. Open the model page to read them.
+
+  01_image.jpeg  (image, 2688x3456)
+  02_image.jpeg  (image, 1664x2944)
+  03_image.jpeg  (image, 1152x1018)
+  04_image.jpeg  (image, 832x1216)
+  05_image.jpeg  (image, 832x1216)
+  06_image.jpeg  (image, 832x1216)
+
+───────────────────────────────────────────────────────────────────────────────
+  HOW TO USE (COMFYUI)
+───────────────────────────────────────────────────────────────────────────────
+  1. Diffusion model -> ComfyUI/models/diffusion_models/. Distilled:
+  flux-2-klein-9b-fp8.safetensors. Base: flux-2-klein-
+  base-9b-fp8.safetensors. (BF16 single-file weights
+  flux-2-klein-9b.safetensors / flux-2-klein-base-9b.safetensors from the
+  gated HF repos also work if you have ~29GB; community GGUF builds exist
+  from unsloth and leejet for lower-VRAM setups.)
+  2. Text encoder ->
+  ComfyUI/models/text_encoders/qwen_3_8b_fp8mixed.safetensors (from Comfy-
+  Org/vae-text-encorder-for-flux-klein-9b). This is the 9B family's encoder;
+  the 4B family uses qwen_3_4b.safetensors — do not mix them.
+  3. VAE -> ComfyUI/models/vae/flux2-vae.safetensors (shared across klein
+  sizes).
+  4. LoRA -> ComfyUI/models/loras/ as a plain .safetensors.
+  5. Load with Load Diffusion Model (UNETLoader) for the transformer,
+  CLIPLoader for the Qwen3 text encoder, VAELoader for the VAE.
+  docs.comfy.org ships prebuilt Flux.2 klein graphs under Workflow Templates
+  for both the distilled and base branches — start from those rather than
+  wiring by hand, since neither the docs nor BFL's repo state
+  sampler/scheduler/shift in prose; those values live in the template.
+  6. Insert LoraLoaderModelOnly between the diffusion-model loader and the
+  sampler. Model-only is correct for klein: the Qwen3 text encoder is not
+  trained by the standard klein trainers, so there is no text-encoder half to
+  patch.
+  7. Set steps and guidance to match the checkpoint you actually loaded, not
+  the one the LoRA was trained on: 4 steps / guidance 1.0 on distilled,
+  ~20-50 steps / guidance ~4.0 on base.
+  8. Note the 9B repos are gated on HuggingFace — you must accept the FLUX
+  Non-Commercial License on the repo page before any download, including
+  scripted ones.
+
+───────────────────────────────────────────────────────────────────────────────
+  Scraped by lorakit on 2026-09-05 03:55 UTC.
+  Stats and description are the author's; pros/cons are derived from the
+  evidence tagged beside each line.
+───────────────────────────────────────────────────────────────────────────────

cg-texture-master-iray-skin-rendering-cg-3s/metadata.json

@@ -0,0 +1,139 @@
+{
+  "name": "CG Texture Master & Iray Skin Rendering | 极致CG质感与3S皮肤材质",
+  "slug": "cg-texture-master-iray-skin-rendering-cg-3s",
+  "base_model": "FLUX.2 [klein] 9B",
+  "media_type": "image",
+  "model_id": 2388820,
+  "version_id": 2686105,
+  "version_name": "Klein9B_V1",
+  "source": "civitai",
+  "url": "https://civitai.com/models/2388820",
+  "repo_id": null,
+  "file_path": null,
+  "author": "nnegret",
+  "category": {
+    "key": "style",
+    "label": "Visual style"
+  },
+  "modalities": [
+    "t2i"
+  ],
+  "primary_modality": "t2i",
+  "checkpoint_family": "klein-9b-distilled",
+  "trigger_words": [
+    "artwork in cgstyle_9b"
+  ],
+  "recommended_settings": {
+    "Sampler": "euler"
+  },
+  "chaining": {
+    "author": {
+      "stacks": "not stated",
+      "notes": "",
+      "avoid": [],
+      "requires": [],
+      "order": "",
+      "strength": "",
+      "stated": false
+    },
+    "general_rules": [
+      "Chain LoRAs by stringing LoraLoaderModelOnly nodes in series, model output into model input. There is no CLIP side to route — klein LoRAs are model-only because the Qwen3 text encoder is not trained (CivitAI's recipe fixes trainTextEncoder: false).",
+      "Only stack LoRAs of the same parameter size. 9B stacks with 9B; a 4B LoRA in a 9B stack is an architectural mismatch (3072 vs 4096 hidden, 5+20 vs 8+24 blocks), not a quality question.",
+      "Mixing a \"Flux.2 Klein 9B\" LoRA and a \"Flux.2 Klein 9B-base\" LoRA in one stack is fine — those strings describe the training checkpoint, not incompatible tensor layouts. Set the sampler for the checkpoint actually loaded.",
+      "Run base-trained LoRAs on the distilled checkpoint when you want speed: HF's klein LoRA guide reports the adapter loads there and typically gives better results than on base, at 4 steps.",
+      "A base-to-turbo / distillation-extraction LoRA is a whole-model delta, not a content LoRA. Put it on the base checkpoint only, and treat its strength as a steps/CFG dial rather than a style knob. Do not add it on top of the already-distilled checkpoint.",
+      "Read each LoRA's own published strength range instead of defaulting everything to 1.0 — klein LoRAs in the wild ship very different ones (0.5 start for the consistency LoRA, 0.8 for SOLRICKS' detail LoRA, under 0.5 and sometimes 0.1 for the vafipas663 distillation delta).",
+      "With a consistency or edit LoRA in the stack, excessive strength suppresses the prompt: the consistency LoRA's article says that if edits stop applying, lower it before changing anything else.",
+      "Prefer runtime LoRA patching over baking a LoRA into the checkpoint: ComfyUI's Model Merge / Model Save nodes have not detected Flux2 Klein models (4B or 9B) since a March 2026 update — issue #13637, still open.",
+      "Step-distilled to exactly 4 inference steps and guidance-distilled; HF card example is num_inference_steps=4, guidance_scale=1.0. Its transformer config has guidance_embeds=false. This is the intended inference target: ComfyUI's blog benchmarks it at ~2s and 19.6GB VRAM on an RTX 5090. Not offered as a training target by BFL's docs or CivitAI's recipe."
+    ]
+  },
+  "pros": [
+    "Well established — 1,579 downloads. [source: CivitAI stats]",
+    "Positively received — 136 thumbs-up with no down-votes. [source: CivitAI stats]",
+    "Explicit trigger word(s) — artwork in cgstyle_9b — so the effect can be turned on and off from the prompt. [source: model metadata]",
+    "Compact at 158 MB — cheap to stack with other LoRAs. [source: file size]",
+    "Author published concrete recommended settings. [source: description]"
+  ],
+  "cons": [],
+  "file": {
+    "name": "nnegrat_cgstyle_9b_1800.safetensors",
+    "size_bytes": 165704168,
+    "size_mb": 158.03,
+    "sha256": "d58f7c9ffb43e3c3cd9f72adbec905f06397218fd885aa86e4e90e96def8b980"
+  },
+  "stats": {
+    "downloads": 1579,
+    "thumbs_up": 136,
+    "thumbs_down": 0
+  },
+  "popularity": {
+    "downloads": 1579,
+    "thumbs_up": 136,
+    "thumbs_down": 0,
+    "comments": 0,
+    "approval": "100% positive (136/136)",
+    "age_days": 203.6,
+    "downloads_per_day": 4.0,
+    "rank_by_downloads": 141,
+    "rank_by_momentum": 141,
+    "of_total": 200,
+    "adoption_tier": "Modest",
+    "momentum": "Long tail"
+  },
+  "selection_score": 70.9,
+  "selection_reasons": [
+    "1,579 downloads",
+    "136 thumbs-up",
+    "documented triggers + usage notes",
+    "top-ranked for its category"
+  ],
+  "example_media": [
+    {
+      "file": "01_image.jpeg",
+      "type": "image",
+      "resolution": "2688x3456",
+      "prompt": ""
+    },
+    {
+      "file": "02_image.jpeg",
+      "type": "image",
+      "resolution": "1664x2944",
+      "prompt": ""
+    },
+    {
+      "file": "03_image.jpeg",
+      "type": "image",
+      "resolution": "1152x1018",
+      "prompt": ""
+    },
+    {
+      "file": "04_image.jpeg",
+      "type": "image",
+      "resolution": "832x1216",
+      "prompt": ""
+    },
+    {
+      "file": "05_image.jpeg",
+      "type": "image",
+      "resolution": "832x1216",
+      "prompt": ""
+    },
+    {
+      "file": "06_image.jpeg",
+      "type": "image",
+      "resolution": "832x1216",
+      "prompt": ""
+    }
+  ],
+  "tags": [
+    "1 girl",
+    "realistic",
+    "anime",
+    "2.5d",
+    "cg",
+    "style"
+  ],
+  "scraped_at": "2026-09-05T03:55:12+00:00",
+  "path": "flux2-klein-9b/cg-texture-master-iray-skin-rendering-cg-3s"
+}

cg-texture-master-iray-skin-rendering-cg-3s/nnegrat_cgstyle_9b_1800.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d58f7c9ffb43e3c3cd9f72adbec905f06397218fd885aa86e4e90e96def8b980
+size 165704168

claire-redfield-resident-evil-2-remake-in-real-life/claireredfieldtailor_flux2-klein_v1a.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a08cadb74bdf244d11d647ee802f969b34f56631dc3d0cf5d7fac19de26f74a4
+size 165704384

claire-redfield-resident-evil-2-remake-in-real-life/info.txt

@@ -0,0 +1,295 @@
+═══════════════════════════════════════════════════════════════════════════════
+  CLAIRE REDFIELD - RESIDENT EVIL 2 REMAKE IN REAL-LIFE
+═══════════════════════════════════════════════════════════════════════════════
+
+───────────────────────────────────────────────────────────────────────────────
+  IDENTITY
+───────────────────────────────────────────────────────────────────────────────
+  Base model        FLUX.2 [klein] 9B
+  Generation mode   not declared by the author
+  Checkpoint        not determined — verify before use
+  Category          Character / subject
+  Author            SteveTailor
+  Source            https://civitai.com/models/2850469
+  Provenance        CivitAI · model 2850469 · version 3227435
+
+───────────────────────────────────────────────────────────────────────────────
+  WEIGHTS
+───────────────────────────────────────────────────────────────────────────────
+  File              claireredfieldtailor_flux2-klein_v1a.safetensors
+  Version           Flux.2 Klein 9B - v1.0A
+  Size              158.0 MB
+  Format            SafeTensor / unknown precision
+  SHA256            a08cadb74bdf244d11d647ee802f969b34f56631dc3d0cf5d7fac19de26f74a4
+  Published         2026-08-15
+
+───────────────────────────────────────────────────────────────────────────────
+  TRIGGER WORDS
+───────────────────────────────────────────────────────────────────────────────
+  › "claireredfieldtailor"
+
+───────────────────────────────────────────────────────────────────────────────
+  RECOMMENDED SETTINGS
+───────────────────────────────────────────────────────────────────────────────
+  Taken verbatim from the author's description:
+
+  LoRA strength       1.0
+  Sampler steps       9
+  CFG / guidance      1.0
+  Sampler             res_multistep
+
+───────────────────────────────────────────────────────────────────────────────
+  ADOPTION & POPULARITY
+───────────────────────────────────────────────────────────────────────────────
+  Downloads               228
+  Rating                  100% positive (21/21)
+  Comments                0
+  Published               21 days ago
+  Download rate           2.7/day since release
+
+  Adoption rank           #197 of 200 in this collection  (Niche)
+  Momentum rank           #183 of 200 by download rate  (Long tail · fresh release)
+
+  Adoption tier is the percentile of total downloads within this collection;
+  momentum is the percentile of downloads-per-day since release. Momentum is
+  ranked rather than measured against a fixed rate, because this base model
+  is itself new and every LoRA looks fast on an absolute scale. Both numbers
+  are CivitAI's, read on the scrape date at the foot of this file.
+
+───────────────────────────────────────────────────────────────────────────────
+  CHAINING / STACKING
+───────────────────────────────────────────────────────────────────────────────
+  Stacks with others      not stated
+
+  The author gave no chaining guidance. The rules below are the
+  general FLUX.2 [klein] 9B ones, not anything specific to this LoRA.
+
+  General FLUX.2 [klein] 9B rules that apply here (see REFERENCE.txt):
+
+  • Chain LoRAs by stringing LoraLoaderModelOnly nodes in series, model
+    output into model input. There is no CLIP side to route — klein LoRAs
+    are model-only because the Qwen3 text encoder is not trained (CivitAI's
+    recipe fixes trainTextEncoder: false).
+  • Only stack LoRAs of the same parameter size. 9B stacks with 9B; a 4B
+    LoRA in a 9B stack is an architectural mismatch (3072 vs 4096 hidden,
+    5+20 vs 8+24 blocks), not a quality question.
+  • Mixing a "Flux.2 Klein 9B" LoRA and a "Flux.2 Klein 9B-base" LoRA in
+    one stack is fine — those strings describe the training checkpoint, not
+    incompatible tensor layouts. Set the sampler for the checkpoint
+    actually loaded.
+  • Run base-trained LoRAs on the distilled checkpoint when you want speed:
+    HF's klein LoRA guide reports the adapter loads there and typically
+    gives better results than on base, at 4 steps.
+  • A base-to-turbo / distillation-extraction LoRA is a whole-model delta,
+    not a content LoRA. Put it on the base checkpoint only, and treat its
+    strength as a steps/CFG dial rather than a style knob. Do not add it on
+    top of the already-distilled checkpoint.
+  • Read each LoRA's own published strength range instead of defaulting
+    everything to 1.0 — klein LoRAs in the wild ship very different ones
+    (0.5 start for the consistency LoRA, 0.8 for SOLRICKS' detail LoRA,
+    under 0.5 and sometimes 0.1 for the vafipas663 distillation delta).
+  • With a consistency or edit LoRA in the stack, excessive strength
+    suppresses the prompt: the consistency LoRA's article says that if
+    edits stop applying, lower it before changing anything else.
+  • Prefer runtime LoRA patching over baking a LoRA into the checkpoint:
+    ComfyUI's Model Merge / Model Save nodes have not detected Flux2 Klein
+    models (4B or 9B) since a March 2026 update — issue #13637, still open.
+  • The author never states which weight family this targets. Two different
+    questions, two different answers. Within 9B (distilled vs base — the
+    two CivitAI strings): not a real incompatibility. Both load through the
+    same Flux2KleinPipeline and have the same repo layout; for the 4B pair,
+    where both transformer configs are public, the distil
+
+───────────────────────────────────────────────────────────────────────────────
+  PROS
+───────────────────────────────────────────────────────────────────────────────
+  + Explicit trigger word(s) — claireredfieldtailor — so the effect can be
+    turned on and off from the prompt. [source: model metadata]
+  + Compact at 158 MB — cheap to stack with other LoRAs. [source: file
+    size]
+  + Author published concrete recommended settings. [source: description]
+
+───────────────────────────────────────────────────────────────────────────────
+  CONS & CAVEATS
+───────────────────────────────────────────────────────────────────────────────
+  - Author never states which generation mode it was trained for; test
+    against your own workflow before relying on it. [source: description]
+  - Modest adoption (228 downloads) — less community feedback to rely on.
+    [source: source stats]
+
+───────────────────────────────────────────────────────────────────────────────
+  COMPATIBILITY WARNING
+───────────────────────────────────────────────────────────────────────────────
+  Two different questions, two different answers. Within 9B (distilled vs
+  base — the two CivitAI strings): not a real incompatibility. Both load
+  through the same Flux2KleinPipeline and have the same repo layout; for the
+  4B pair, where both transformer configs are public, the distilled and base
+  configs are byte-for-byte identical in every shape field, and the 9B pair
+  is built the same way. HF's own klein LoRA guide states the workflow
+  directly: "Distilled models are step-compressed for fast inference; you
+  train against the base checkpoint and the adapter still loads on the
+  distilled model afterward", and "Applying the LoRA on the distilled model
+  typically gives better results than the base model, and it's faster" (that
+  post is written against 4B, but the family structure is the same). So the
+  two CivitAI strings record which checkpoint the author trained or previewed
+  against, not a hard compatibility wall. The real risk is a silent settings
+  mismatch: load a base-trained LoRA on the distilled checkpoint while
+  keeping the author's 20-50 step / CFG ~4 preview settings and the output
+  degrades badly (distilled wants 4 steps / guidance 1.0), or the reverse.
+  Nothing errors — it just looks wrong. Across sizes (9B vs 4B): a genuine
+  architectural mismatch, from the published transformer configs — different
+  hidden width (4096 vs 3072), different block counts (8+24 vs 5+20) and a
+  different text-conditioning dim (12288 vs 7680, following Qwen3-8B vs
+  Qwen3-4B). A 4B LoRA cannot patch a 9B model: expect missing keys and shape
+  errors, not degraded output. The exact ComfyUI error text is unverified. KV
+  variant: same 9B size and same 4-step distillation, but a distinct
+  checkpoint with its own pipeline; whether standard klein 9B LoRAs patch it
+  cleanly is not documented — see unverified.
+
+───────────────────────────────────────────────────────────────────────────────
+  AUTHOR'S DESCRIPTION
+───────────────────────────────────────────────────────────────────────────────
+  Meet Claire Redfield — the one who rode into Raccoon City looking for her
+  brother and never quite stopped moving since. This LoRA is a fan tribute
+  built around her likeness: her face, that level, unimpressed calm she keeps
+  even when everything around her is on fire, and the warmth underneath it.
+  One thing up front, because it changes how you prompt her: this is a
+  character-identity LoRA. It reproduces who she is, not a fixed costume. The
+  biker jacket, the red vest, the Raccoon City era gear — none of that is
+  baked in. The dataset is deliberately varied everyday clothing and
+  locations, so her identity holds up in any setting. If you want the
+  signature look, describe it in your prompt; the LoRA keeps her face while
+  you decide the wardrobe.
+  It is tuned for Z-Image Turbo , so you get a photorealistic Claire at turbo
+  speed: few steps, low CFG, fast iterations, strong identity. There are also
+  versions for Illustrious and FLUX.2 Klein 9B . Have fun with her — and
+  please share what you make.
+  Recommended Settings - Trigger word: claireredfieldtailor (put it at the
+  very start of your prompt)
+
+  - Strength: 1.0 (useful range 0.8 – 1.0)
+
+  Z-Image-Turbo - Steps: 9
+
+  - CFG: 1.0
+
+  - Sampler: res_multistep
+
+  - Scheduler: simple
+
+  - Shift: 3.0 (via a ModelSamplingAuraFlow node — Z-Image / AuraFlow
+  specific)
+
+  - Negative prompt: (empty)
+
+  - Base model: Z-Image Turbo. A realistic Z-Image Turbo checkpoint such as
+  AnalogMadness (Z-Image Turbo) v2.0 or cyberrealistic Z-Image v0.5 gives the
+  most photographic look.
+
+  Illustrious - Steps: 30
+
+  - CFG: 4.0-5.0
+
+  - Sampler: dpmpp_2m, dpmpp_3m_sde
+
+  - Scheduler: karras
+
+  - Models: Realism Illustrious By Stable Yogi, Ultra Realistic By Stable
+  Yogi, ramthrustsNSFWPINK_alchemyMix176, perfectdeliberate v70, iLustMix
+
+  Flux.2 Klein 9B - Steps: 25-30
+
+  - CFG: 1.0
+
+  - Guidance: 3.0-4.0
+
+  - Sampler: euler / res_multistep
+
+  - Scheduler: simple
+
+  - Negative prompt: (empty)
+
+  - Base model: Flux.2 Klein 9B + FP8, Jib Mix Klein,
+
+  Prompt Examples
+  Prompt Examples Portrait prompt: A photo of claireredfieldtailor, a woman.
+  A half-body portrait, three-quarter view, soft window light, warm tones,
+  shallow depth of field, photorealistic, professional photography. Meaning:
+  just her, cleanly lit, nothing competing with the face. Use this to verify
+  the likeness
+  before you build a bigger scene.
+
+  Training Details - Source: curated character dataset (portraits, full-body,
+  scenes, outside, inside, ...)
+
+  - Dataset size: 89 images
+
+  - Resolution: 768 px
+
+  Notes - Fan service / disclaimer: This is a fan-made tribute LoRA. It is
+  not affiliated with, endorsed by, or connected to Capcom or any Resident
+  Evil rights holder. Claire Redfield is a fictional character owned by
+  Capcom; this LoRA reproduces the likeness of a fictional character, not of
+  a real person. Check Civitai's rules and Capcom's position before enabling
+  commercial use or on-site generation. For personal, non-commercial fan use.
+
+  - Identity, not a costume: the biker jacket, the red vest and the rest of
+  the game wardrobe were not in the training data. Describe the clothing you
+  want in your prompt — the LoRA supplies the face. Any accessory has to be
+  prompted and only makes sense in front or side views.
+
+───────────────────────────────────────────────────────────────────────────────
+  EXAMPLE MEDIA
+───────────────────────────────────────────────────────────────────────────────
+  6 preview(s) in ./example_images/ — community results for
+  this LoRA. Videos keep a matching _poster.jpg still frame.
+
+  No prompts are recorded below: CivitAI's public API flags these posts as
+  having generation metadata but does not return it, so the prompts cannot be
+  scraped. Open the model page to read them.
+
+  01_image.jpeg  (image, 896x1152)
+  02_image.jpeg  (image, 896x1152)
+  03_image.jpeg  (image, 896x1152)
+  04_image.jpeg  (image, 896x1152)
+  05_image.jpeg  (image, 896x1152)
+  06_image.jpeg  (image, 832x1216)
+
+───────────────────────────────────────────────────────────────────────────────
+  HOW TO USE (COMFYUI)
+───────────────────────────────────────────────────────────────────────────────
+  1. Diffusion model -> ComfyUI/models/diffusion_models/. Distilled:
+  flux-2-klein-9b-fp8.safetensors. Base: flux-2-klein-
+  base-9b-fp8.safetensors. (BF16 single-file weights
+  flux-2-klein-9b.safetensors / flux-2-klein-base-9b.safetensors from the
+  gated HF repos also work if you have ~29GB; community GGUF builds exist
+  from unsloth and leejet for lower-VRAM setups.)
+  2. Text encoder ->
+  ComfyUI/models/text_encoders/qwen_3_8b_fp8mixed.safetensors (from Comfy-
+  Org/vae-text-encorder-for-flux-klein-9b). This is the 9B family's encoder;
+  the 4B family uses qwen_3_4b.safetensors — do not mix them.
+  3. VAE -> ComfyUI/models/vae/flux2-vae.safetensors (shared across klein
+  sizes).
+  4. LoRA -> ComfyUI/models/loras/ as a plain .safetensors.
+  5. Load with Load Diffusion Model (UNETLoader) for the transformer,
+  CLIPLoader for the Qwen3 text encoder, VAELoader for the VAE.
+  docs.comfy.org ships prebuilt Flux.2 klein graphs under Workflow Templates
+  for both the distilled and base branches — start from those rather than
+  wiring by hand, since neither the docs nor BFL's repo state
+  sampler/scheduler/shift in prose; those values live in the template.
+  6. Insert LoraLoaderModelOnly between the diffusion-model loader and the
+  sampler. Model-only is correct for klein: the Qwen3 text encoder is not
+  trained by the standard klein trainers, so there is no text-encoder half to
+  patch.
+  7. Set steps and guidance to match the checkpoint you actually loaded, not
+  the one the LoRA was trained on: 4 steps / guidance 1.0 on distilled,
+  ~20-50 steps / guidance ~4.0 on base.
+  8. Note the 9B repos are gated on HuggingFace — you must accept the FLUX
+  Non-Commercial License on the repo page before any download, including
+  scripted ones.
+
+───────────────────────────────────────────────────────────────────────────────
+  Scraped by lorakit on 2026-09-05 03:55 UTC.
+  Stats and description are the author's; pros/cons are derived from the
+  evidence tagged beside each line.
+───────────────────────────────────────────────────────────────────────────────

claire-redfield-resident-evil-2-remake-in-real-life/metadata.json

@@ -0,0 +1,142 @@
+{
+  "name": "Claire Redfield - Resident Evil 2 Remake in Real-Life",
+  "slug": "claire-redfield-resident-evil-2-remake-in-real-life",
+  "base_model": "FLUX.2 [klein] 9B",
+  "media_type": "image",
+  "model_id": 2850469,
+  "version_id": 3227435,
+  "version_name": "Flux.2 Klein 9B - v1.0A",
+  "source": "civitai",
+  "url": "https://civitai.com/models/2850469",
+  "repo_id": null,
+  "file_path": null,
+  "author": "SteveTailor",
+  "category": {
+    "key": "character",
+    "label": "Character / subject"
+  },
+  "modalities": [],
+  "primary_modality": "unspecified",
+  "checkpoint_family": "unknown",
+  "trigger_words": [
+    "claireredfieldtailor"
+  ],
+  "recommended_settings": {
+    "LoRA strength": "1.0",
+    "Sampler steps": "9",
+    "CFG / guidance": "1.0",
+    "Sampler": "res_multistep"
+  },
+  "chaining": {
+    "author": {
+      "stacks": "not stated",
+      "notes": "",
+      "avoid": [],
+      "requires": [],
+      "order": "",
+      "strength": "",
+      "stated": false
+    },
+    "general_rules": [
+      "Chain LoRAs by stringing LoraLoaderModelOnly nodes in series, model output into model input. There is no CLIP side to route — klein LoRAs are model-only because the Qwen3 text encoder is not trained (CivitAI's recipe fixes trainTextEncoder: false).",
+      "Only stack LoRAs of the same parameter size. 9B stacks with 9B; a 4B LoRA in a 9B stack is an architectural mismatch (3072 vs 4096 hidden, 5+20 vs 8+24 blocks), not a quality question.",
+      "Mixing a \"Flux.2 Klein 9B\" LoRA and a \"Flux.2 Klein 9B-base\" LoRA in one stack is fine — those strings describe the training checkpoint, not incompatible tensor layouts. Set the sampler for the checkpoint actually loaded.",
+      "Run base-trained LoRAs on the distilled checkpoint when you want speed: HF's klein LoRA guide reports the adapter loads there and typically gives better results than on base, at 4 steps.",
+      "A base-to-turbo / distillation-extraction LoRA is a whole-model delta, not a content LoRA. Put it on the base checkpoint only, and treat its strength as a steps/CFG dial rather than a style knob. Do not add it on top of the already-distilled checkpoint.",
+      "Read each LoRA's own published strength range instead of defaulting everything to 1.0 — klein LoRAs in the wild ship very different ones (0.5 start for the consistency LoRA, 0.8 for SOLRICKS' detail LoRA, under 0.5 and sometimes 0.1 for the vafipas663 distillation delta).",
+      "With a consistency or edit LoRA in the stack, excessive strength suppresses the prompt: the consistency LoRA's article says that if edits stop applying, lower it before changing anything else.",
+      "Prefer runtime LoRA patching over baking a LoRA into the checkpoint: ComfyUI's Model Merge / Model Save nodes have not detected Flux2 Klein models (4B or 9B) since a March 2026 update — issue #13637, still open.",
+      "The author never states which weight family this targets. Two different questions, two different answers. Within 9B (distilled vs base — the two CivitAI strings): not a real incompatibility. Both load through the same Flux2KleinPipeline and have the same repo layout; for the 4B pair, where both transformer configs are public, the distil"
+    ]
+  },
+  "pros": [
+    "Explicit trigger word(s) — claireredfieldtailor — so the effect can be turned on and off from the prompt. [source: model metadata]",
+    "Compact at 158 MB — cheap to stack with other LoRAs. [source: file size]",
+    "Author published concrete recommended settings. [source: description]"
+  ],
+  "cons": [
+    "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]",
+    "Modest adoption (228 downloads) — less community feedback to rely on. [source: source stats]"
+  ],
+  "file": {
+    "name": "claireredfieldtailor_flux2-klein_v1a.safetensors",
+    "size_bytes": 165704384,
+    "size_mb": 158.03,
+    "sha256": "a08cadb74bdf244d11d647ee802f969b34f56631dc3d0cf5d7fac19de26f74a4"
+  },
+  "stats": {
+    "downloads": 228,
+    "thumbs_up": 21,
+    "thumbs_down": 0
+  },
+  "popularity": {
+    "downloads": 228,
+    "thumbs_up": 21,
+    "thumbs_down": 0,
+    "comments": 0,
+    "approval": "100% positive (21/21)",
+    "age_days": 20.8,
+    "downloads_per_day": 2.69,
+    "rank_by_downloads": 197,
+    "rank_by_momentum": 183,
+    "of_total": 200,
+    "adoption_tier": "Niche",
+    "momentum": "Long tail · fresh release"
+  },
+  "selection_score": 70.9,
+  "selection_reasons": [
+    "228 downloads",
+    "21 thumbs-up",
+    "documented triggers + usage notes",
+    "top-ranked for its category"
+  ],
+  "example_media": [
+    {
+      "file": "01_image.jpeg",
+      "type": "image",
+      "resolution": "896x1152",
+      "prompt": ""
+    },
+    {
+      "file": "02_image.jpeg",
+      "type": "image",
+      "resolution": "896x1152",
+      "prompt": ""
+    },
+    {
+      "file": "03_image.jpeg",
+      "type": "image",
+      "resolution": "896x1152",
+      "prompt": ""
+    },
+    {
+      "file": "04_image.jpeg",
+      "type": "image",
+      "resolution": "896x1152",
+      "prompt": ""
+    },
+    {
+      "file": "05_image.jpeg",
+      "type": "image",
+      "resolution": "896x1152",
+      "prompt": ""
+    },
+    {
+      "file": "06_image.jpeg",
+      "type": "image",
+      "resolution": "832x1216",
+      "prompt": ""
+    }
+  ],
+  "tags": [
+    "claire redfield",
+    "character",
+    "resi2",
+    "resident evil 2 remake",
+    "resident evil 2",
+    "claire",
+    "redfield"
+  ],
+  "scraped_at": "2026-09-05T03:55:48+00:00",
+  "path": "flux2-klein-9b/claire-redfield-resident-evil-2-remake-in-real-life"
+}

detailer-ce/Detailer04d_CE_Flux2K9b_AIT_000000150.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebd9c7088864987efc067f6925dce369c8fe6964137aff044a5d4c8b6650bfd9
+size 41447688

detailer-ce/info.txt

@@ -0,0 +1,238 @@
+═══════════════════════════════════════════════════════════════════════════════
+  DETAILER CE
+═══════════════════════════════════════════════════════════════════════════════
+
+───────────────────────────────────────────────────────────────────────────────
+  IDENTITY
+───────────────────────────────────────────────────────────────────────────────
+  Base model        FLUX.2 [klein] 9B
+  Generation mode   not declared by the author
+  Checkpoint        not determined — verify before use
+  Category          Detail / anatomy fixer
+  Author            CreativeEdge
+  Source            https://civitai.com/models/2574468
+  Provenance        CivitAI · model 2574468 · version 2896490
+
+───────────────────────────────────────────────────────────────────────────────
+  WEIGHTS
+───────────────────────────────────────────────────────────────────────────────
+  File              Detailer04d_CE_Flux2K9b_AIT_000000150.safetensors
+  Version           V04d - Flux.2 Klein 9b
+  Size              39.5 MB
+  Format            SafeTensor / unknown precision
+  SHA256            ebd9c7088864987efc067f6925dce369c8fe6964137aff044a5d4c8b6650bfd9
+  Published         2026-04-28
+
+───────────────────────────────────────────────────────────────────────────────
+  TRIGGER WORDS
+───────────────────────────────────────────────────────────────────────────────
+  No trigger word. The LoRA applies as soon as it is loaded —
+  strength is the only way to dial the effect up or down.
+
+───────────────────────────────────────────────────────────────────────────────
+  RECOMMENDED SETTINGS
+───────────────────────────────────────────────────────────────────────────────
+  The author did not publish specific settings.
+  No author settings were published. Community baseline: steps Distilled
+  FLUX.2-klein-9B: 4 (HF card example; it is step-distilled to exactly this,
+  and more steps do not refine further). Base FLUX.2-klein-base-9B: 50 in the
+  HF card example and in ComfyUI's benchmark; community practice runs lower,
+  with ~20 for production and 8-12 for fast iteration reported in the CivitAI
+  consistency-LoRA article., cfg Distilled: guidance_scale 1.0 (HF card). It
+  is guidance-distilled and its transformer config sets
+  guidance_embeds=false, so raising CFG degrades rather than improves. Base:
+  guidance_scale 4.0 (HF card); the CivitAI consistency-LoRA article
+  recommends 3.5-4.5 for most edits., sampler Not stated by BFL or by
+  docs.comfy.org, and the official flux2 GitHub repo documents no sampler,
+  scheduler or shift values either. Community consensus is euler with a
+  resolution-dependent shifted-Euler / flow-match schedule (ComfyUI's Flux2
+  workflows use a scheduler that interpolates shift between base_shift and
+  max_shift by output resolution). Start from ComfyUI's downloadable Flux.2
+  klein workflow template rather than a blog. See unverified., resolution
+  1024x1024 in every BFL HF card example and in CivitAI's 9B training recipe
+  ("resolution": 1024). Non-square ratios work but no official list of
+  supported dimensions or a stated alignment multiple was found — see
+  unverified., lora_strength 1.0 is the diffusers/ComfyUI default and what
+  CivitAI applies to training previews (strength 1.0). Published per-LoRA
+  guidance varies and should be read from the LoRA, not assumed: the widely
+  used Klein 9B consistency LoRA's CivitAI article says start at 0.5 within a
+  0.2-1.0 range (0.2-0.4 loose, 0.5-0.7 balanced, 0.8-1.0 maximum structural
+  stability) and warns that too high a strength stops edits from applying;
+  SOLRICKS' Realistic Detail LoRA card says 0.6-0.8 subtle, 0.8-1.0 stronger,
+  start 0.8; the vafipas663 distillation-delta LoRA card warns it "produces
+  static noise at strength > 0.5. And sometimes, you have to go as low as
+  0.1.".
+
+───────────────────────────────────────────────────────────────────────────────
+  ADOPTION & POPULARITY
+───────────────────────────────────────────────────────────────────────────────
+  Downloads               1,961
+  Rating                  100% positive (127/127)
+  Comments                0
+  Published               129 days ago
+  Download rate           4.1/day since release
+
+  Adoption rank           #133 of 200 in this collection  (Modest)
+  Momentum rank           #136 of 200 by download rate  (Long tail)
+
+  Adoption tier is the percentile of total downloads within this collection;
+  momentum is the percentile of downloads-per-day since release. Momentum is
+  ranked rather than measured against a fixed rate, because this base model
+  is itself new and every LoRA looks fast on an absolute scale. Both numbers
+  are CivitAI's, read on the scrape date at the foot of this file.
+
+───────────────────────────────────────────────────────────────────────────────
+  CHAINING / STACKING
+───────────────────────────────────────────────────────────────────────────────
+  Stacks with others      not stated
+
+  The author gave no chaining guidance. The rules below are the
+  general FLUX.2 [klein] 9B ones, not anything specific to this LoRA.
+
+  General FLUX.2 [klein] 9B rules that apply here (see REFERENCE.txt):
+
+  • Chain LoRAs by stringing LoraLoaderModelOnly nodes in series, model
+    output into model input. There is no CLIP side to route — klein LoRAs
+    are model-only because the Qwen3 text encoder is not trained (CivitAI's
+    recipe fixes trainTextEncoder: false).
+  • Only stack LoRAs of the same parameter size. 9B stacks with 9B; a 4B
+    LoRA in a 9B stack is an architectural mismatch (3072 vs 4096 hidden,
+    5+20 vs 8+24 blocks), not a quality question.
+  • Mixing a "Flux.2 Klein 9B" LoRA and a "Flux.2 Klein 9B-base" LoRA in
+    one stack is fine — those strings describe the training checkpoint, not
+    incompatible tensor layouts. Set the sampler for the checkpoint
+    actually loaded.
+  • Run base-trained LoRAs on the distilled checkpoint when you want speed:
+    HF's klein LoRA guide reports the adapter loads there and typically
+    gives better results than on base, at 4 steps.
+  • A base-to-turbo / distillation-extraction LoRA is a whole-model delta,
+    not a content LoRA. Put it on the base checkpoint only, and treat its
+    strength as a steps/CFG dial rather than a style knob. Do not add it on
+    top of the already-distilled checkpoint.
+  • Read each LoRA's own published strength range instead of defaulting
+    everything to 1.0 — klein LoRAs in the wild ship very different ones
+    (0.5 start for the consistency LoRA, 0.8 for SOLRICKS' detail LoRA,
+    under 0.5 and sometimes 0.1 for the vafipas663 distillation delta).
+  • With a consistency or edit LoRA in the stack, excessive strength
+    suppresses the prompt: the consistency LoRA's article says that if
+    edits stop applying, lower it before changing anything else.
+  • Prefer runtime LoRA patching over baking a LoRA into the checkpoint:
+    ComfyUI's Model Merge / Model Save nodes have not detected Flux2 Klein
+    models (4B or 9B) since a March 2026 update — issue #13637, still open.
+  • The author never states which weight family this targets. Two different
+    questions, two different answers. Within 9B (distilled vs base — the
+    two CivitAI strings): not a real incompatibility. Both load through the
+    same Flux2KleinPipeline and have the same repo layout; for the 4B pair,
+    where both transformer configs are public, the distil
+
+───────────────────────────────────────────────────────────────────────────────
+  PROS
+───────────────────────────────────────────────────────────────────────────────
+  + Well established — 1,961 downloads. [source: CivitAI stats]
+  + Positively received — 127 thumbs-up with no down-votes. [source:
+    CivitAI stats]
+  + Compact at 40 MB — cheap to stack with other LoRAs. [source: file size]
+
+───────────────────────────────────────────────────────────────────────────────
+  CONS & CAVEATS
+───────────────────────────────────────────────────────────────────────────────
+  - Needs a reduced strength; the default 1.0 is likely too strong.
+    [source: description]
+  - No trigger word — the LoRA is always on once loaded, so strength is the
+    only control. [source: model metadata]
+  - Author never states which generation mode it was trained for; test
+    against your own workflow before relying on it. [source: description]
+
+───────────────────────────────────────────────────────────────────────────────
+  COMPATIBILITY WARNING
+───────────────────────────────────────────────────────────────────────────────
+  Two different questions, two different answers. Within 9B (distilled vs
+  base — the two CivitAI strings): not a real incompatibility. Both load
+  through the same Flux2KleinPipeline and have the same repo layout; for the
+  4B pair, where both transformer configs are public, the distilled and base
+  configs are byte-for-byte identical in every shape field, and the 9B pair
+  is built the same way. HF's own klein LoRA guide states the workflow
+  directly: "Distilled models are step-compressed for fast inference; you
+  train against the base checkpoint and the adapter still loads on the
+  distilled model afterward", and "Applying the LoRA on the distilled model
+  typically gives better results than the base model, and it's faster" (that
+  post is written against 4B, but the family structure is the same). So the
+  two CivitAI strings record which checkpoint the author trained or previewed
+  against, not a hard compatibility wall. The real risk is a silent settings
+  mismatch: load a base-trained LoRA on the distilled checkpoint while
+  keeping the author's 20-50 step / CFG ~4 preview settings and the output
+  degrades badly (distilled wants 4 steps / guidance 1.0), or the reverse.
+  Nothing errors — it just looks wrong. Across sizes (9B vs 4B): a genuine
+  architectural mismatch, from the published transformer configs — different
+  hidden width (4096 vs 3072), different block counts (8+24 vs 5+20) and a
+  different text-conditioning dim (12288 vs 7680, following Qwen3-8B vs
+  Qwen3-4B). A 4B LoRA cannot patch a 9B model: expect missing keys and shape
+  errors, not degraded output. The exact ComfyUI error text is unverified. KV
+  variant: same 9B size and same 4-step distillation, but a distinct
+  checkpoint with its own pipeline; whether standard klein 9B LoRAs patch it
+  cleanly is not documented — see unverified.
+
+───────────────────────────────────────────────────────────────────────────────
+  AUTHOR'S DESCRIPTION
+───────────────────────────────────────────────────────────────────────────────
+  Detail slider. On its own, the useful range is anywhere from around -1 to
+  positive 1, though some of the models top out around +/- 0.8. See the
+  showcase images for an idea of the results at various settings. When used
+  in conjunction with other LoRAs, it can be more useful to lower the slider
+  strength to around 0.4.
+
+───────────────────────────────────────────────────────────────────────────────
+  EXAMPLE MEDIA
+───────────────────────────────────────────────────────────────────────────────
+  6 preview(s) in ./example_images/ — community results for
+  this LoRA. Videos keep a matching _poster.jpg still frame.
+
+  No prompts are recorded below: CivitAI's public API flags these posts as
+  having generation metadata but does not return it, so the prompts cannot be
+  scraped. Open the model page to read them.
+
+  01_image.jpeg  (image, 3072x1024)
+  02_image.jpeg  (image, 3072x1024)
+  03_image.jpeg  (image, 3072x1024)
+  04_image.jpeg  (image, 1024x1024)
+  05_image.jpeg  (image, 1024x1024)
+  06_image.jpeg  (image, 1024x1024)
+
+───────────────────────────────────────────────────────────────────────────────
+  HOW TO USE (COMFYUI)
+───────────────────────────────────────────────────────────────────────────────
+  1. Diffusion model -> ComfyUI/models/diffusion_models/. Distilled:
+  flux-2-klein-9b-fp8.safetensors. Base: flux-2-klein-
+  base-9b-fp8.safetensors. (BF16 single-file weights
+  flux-2-klein-9b.safetensors / flux-2-klein-base-9b.safetensors from the
+  gated HF repos also work if you have ~29GB; community GGUF builds exist
+  from unsloth and leejet for lower-VRAM setups.)
+  2. Text encoder ->
+  ComfyUI/models/text_encoders/qwen_3_8b_fp8mixed.safetensors (from Comfy-
+  Org/vae-text-encorder-for-flux-klein-9b). This is the 9B family's encoder;
+  the 4B family uses qwen_3_4b.safetensors — do not mix them.
+  3. VAE -> ComfyUI/models/vae/flux2-vae.safetensors (shared across klein
+  sizes).
+  4. LoRA -> ComfyUI/models/loras/ as a plain .safetensors.
+  5. Load with Load Diffusion Model (UNETLoader) for the transformer,
+  CLIPLoader for the Qwen3 text encoder, VAELoader for the VAE.
+  docs.comfy.org ships prebuilt Flux.2 klein graphs under Workflow Templates
+  for both the distilled and base branches — start from those rather than
+  wiring by hand, since neither the docs nor BFL's repo state
+  sampler/scheduler/shift in prose; those values live in the template.
+  6. Insert LoraLoaderModelOnly between the diffusion-model loader and the
+  sampler. Model-only is correct for klein: the Qwen3 text encoder is not
+  trained by the standard klein trainers, so there is no text-encoder half to
+  patch.
+  7. Set steps and guidance to match the checkpoint you actually loaded, not
+  the one the LoRA was trained on: 4 steps / guidance 1.0 on distilled,
+  ~20-50 steps / guidance ~4.0 on base.
+  8. Note the 9B repos are gated on HuggingFace — you must accept the FLUX
+  Non-Commercial License on the repo page before any download, including
+  scripted ones.
+
+───────────────────────────────────────────────────────────────────────────────
+  Scraped by lorakit on 2026-09-05 03:54 UTC.
+  Stats and description are the author's; pros/cons are derived from the
+  evidence tagged beside each line.
+───────────────────────────────────────────────────────────────────────────────

detailer-ce/metadata.json

@@ -0,0 +1,132 @@
+{
+  "name": "Detailer CE",
+  "slug": "detailer-ce",
+  "base_model": "FLUX.2 [klein] 9B",
+  "media_type": "image",
+  "model_id": 2574468,
+  "version_id": 2896490,
+  "version_name": "V04d - Flux.2 Klein 9b",
+  "source": "civitai",
+  "url": "https://civitai.com/models/2574468",
+  "repo_id": null,
+  "file_path": null,
+  "author": "CreativeEdge",
+  "category": {
+    "key": "detail",
+    "label": "Detail / anatomy fixer"
+  },
+  "modalities": [],
+  "primary_modality": "unspecified",
+  "checkpoint_family": "unknown",
+  "trigger_words": [],
+  "recommended_settings": {},
+  "chaining": {
+    "author": {
+      "stacks": "not stated",
+      "notes": "",
+      "avoid": [],
+      "requires": [],
+      "order": "",
+      "strength": "",
+      "stated": false
+    },
+    "general_rules": [
+      "Chain LoRAs by stringing LoraLoaderModelOnly nodes in series, model output into model input. There is no CLIP side to route — klein LoRAs are model-only because the Qwen3 text encoder is not trained (CivitAI's recipe fixes trainTextEncoder: false).",
+      "Only stack LoRAs of the same parameter size. 9B stacks with 9B; a 4B LoRA in a 9B stack is an architectural mismatch (3072 vs 4096 hidden, 5+20 vs 8+24 blocks), not a quality question.",
+      "Mixing a \"Flux.2 Klein 9B\" LoRA and a \"Flux.2 Klein 9B-base\" LoRA in one stack is fine — those strings describe the training checkpoint, not incompatible tensor layouts. Set the sampler for the checkpoint actually loaded.",
+      "Run base-trained LoRAs on the distilled checkpoint when you want speed: HF's klein LoRA guide reports the adapter loads there and typically gives better results than on base, at 4 steps.",
+      "A base-to-turbo / distillation-extraction LoRA is a whole-model delta, not a content LoRA. Put it on the base checkpoint only, and treat its strength as a steps/CFG dial rather than a style knob. Do not add it on top of the already-distilled checkpoint.",
+      "Read each LoRA's own published strength range instead of defaulting everything to 1.0 — klein LoRAs in the wild ship very different ones (0.5 start for the consistency LoRA, 0.8 for SOLRICKS' detail LoRA, under 0.5 and sometimes 0.1 for the vafipas663 distillation delta).",
+      "With a consistency or edit LoRA in the stack, excessive strength suppresses the prompt: the consistency LoRA's article says that if edits stop applying, lower it before changing anything else.",
+      "Prefer runtime LoRA patching over baking a LoRA into the checkpoint: ComfyUI's Model Merge / Model Save nodes have not detected Flux2 Klein models (4B or 9B) since a March 2026 update — issue #13637, still open.",
+      "The author never states which weight family this targets. Two different questions, two different answers. Within 9B (distilled vs base — the two CivitAI strings): not a real incompatibility. Both load through the same Flux2KleinPipeline and have the same repo layout; for the 4B pair, where both transformer configs are public, the distil"
+    ]
+  },
+  "pros": [
+    "Well established — 1,961 downloads. [source: CivitAI stats]",
+    "Positively received — 127 thumbs-up with no down-votes. [source: CivitAI stats]",
+    "Compact at 40 MB — cheap to stack with other LoRAs. [source: file size]"
+  ],
+  "cons": [
+    "Needs a reduced strength; the default 1.0 is likely too strong. [source: description]",
+    "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]",
+    "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]"
+  ],
+  "file": {
+    "name": "Detailer04d_CE_Flux2K9b_AIT_000000150.safetensors",
+    "size_bytes": 41447688,
+    "size_mb": 39.53,
+    "sha256": "ebd9c7088864987efc067f6925dce369c8fe6964137aff044a5d4c8b6650bfd9"
+  },
+  "stats": {
+    "downloads": 1961,
+    "thumbs_up": 127,
+    "thumbs_down": 0
+  },
+  "popularity": {
+    "downloads": 1961,
+    "thumbs_up": 127,
+    "thumbs_down": 0,
+    "comments": 0,
+    "approval": "100% positive (127/127)",
+    "age_days": 129.4,
+    "downloads_per_day": 4.12,
+    "rank_by_downloads": 133,
+    "rank_by_momentum": 136,
+    "of_total": 200,
+    "adoption_tier": "Modest",
+    "momentum": "Long tail"
+  },
+  "selection_score": 71.3,
+  "selection_reasons": [
+    "1,961 downloads",
+    "127 thumbs-up",
+    "top-ranked for its category"
+  ],
+  "example_media": [
+    {
+      "file": "01_image.jpeg",
+      "type": "image",
+      "resolution": "3072x1024",
+      "prompt": ""
+    },
+    {
+      "file": "02_image.jpeg",
+      "type": "image",
+      "resolution": "3072x1024",
+      "prompt": ""
+    },
+    {
+      "file": "03_image.jpeg",
+      "type": "image",
+      "resolution": "3072x1024",
+      "prompt": ""
+    },
+    {
+      "file": "04_image.jpeg",
+      "type": "image",
+      "resolution": "1024x1024",
+      "prompt": ""
+    },
+    {
+      "file": "05_image.jpeg",
+      "type": "image",
+      "resolution": "1024x1024",
+      "prompt": ""
+    },
+    {
+      "file": "06_image.jpeg",
+      "type": "image",
+      "resolution": "1024x1024",
+      "prompt": ""
+    }
+  ],
+  "tags": [
+    "detail slider",
+    "detail",
+    "style",
+    "slider"
+  ],
+  "scraped_at": "2026-09-05T03:54:59+00:00",
+  "path": "flux2-klein-9b/detailer-ce"
+}

hand-drawn-sketch/hand_drawn_sketch_klein9b.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96b1ce30cadd42bc15bb457117814dd0d31e19172e236f3038fc877129e869aa
+size 82879480

hand-drawn-sketch/info.txt

@@ -0,0 +1,217 @@
+═══════════════════════════════════════════════════════════════════════════════
+  HAND-DRAWN SKETCH
+═══════════════════════════════════════════════════════════════════════════════
+
+───────────────────────────────────────────────────────────────────────────────
+  IDENTITY
+───────────────────────────────────────────────────────────────────────────────
+  Base model        FLUX.2 [klein] 9B
+  Generation mode   not declared by the author
+  Checkpoint        not determined — verify before use
+  Category          Visual style
+  Author            Yofaraway
+  Source            https://civitai.com/models/2692445
+  Provenance        CivitAI · model 2692445 · version 3023139
+
+───────────────────────────────────────────────────────────────────────────────
+  WEIGHTS
+───────────────────────────────────────────────────────────────────────────────
+  File              hand_drawn_sketch_klein9b.safetensors
+  Version           Klein9B v1
+  Size              79.0 MB
+  Format            SafeTensor / unknown precision
+  SHA256            96b1ce30cadd42bc15bb457117814dd0d31e19172e236f3038fc877129e869aa
+  Published         2026-06-16
+
+───────────────────────────────────────────────────────────────────────────────
+  TRIGGER WORDS
+───────────────────────────────────────────────────────────────────────────────
+  › "hand-drawn sketch"
+
+───────────────────────────────────────────────────────────────────────────────
+  RECOMMENDED SETTINGS
+───────────────────────────────────────────────────────────────────────────────
+  Taken verbatim from the author's description:
+
+  Sampler steps       4
+  Resolution          768x1152
+  CFG / guidance      1
+  Sampler             Euler
+
+───────────────────────────────────────────────────────────────────────────────
+  ADOPTION & POPULARITY
+───────────────────────────────────────────────────────────────────────────────
+  Downloads               361
+  Rating                  100% positive (48/48)
+  Comments                1
+  Published               81 days ago
+  Download rate           4.5/day since release
+
+  Adoption rank           #192 of 200 in this collection  (Niche)
+  Momentum rank           #127 of 200 by download rate  (Long tail)
+
+  Adoption tier is the percentile of total downloads within this collection;
+  momentum is the percentile of downloads-per-day since release. Momentum is
+  ranked rather than measured against a fixed rate, because this base model
+  is itself new and every LoRA looks fast on an absolute scale. Both numbers
+  are CivitAI's, read on the scrape date at the foot of this file.
+
+───────────────────────────────────────────────────────────────────────────────
+  CHAINING / STACKING
+───────────────────────────────────────────────────────────────────────────────
+  Stacks with others      not stated
+
+  The author gave no chaining guidance. The rules below are the
+  general FLUX.2 [klein] 9B ones, not anything specific to this LoRA.
+
+  General FLUX.2 [klein] 9B rules that apply here (see REFERENCE.txt):
+
+  • Chain LoRAs by stringing LoraLoaderModelOnly nodes in series, model
+    output into model input. There is no CLIP side to route — klein LoRAs
+    are model-only because the Qwen3 text encoder is not trained (CivitAI's
+    recipe fixes trainTextEncoder: false).
+  • Only stack LoRAs of the same parameter size. 9B stacks with 9B; a 4B
+    LoRA in a 9B stack is an architectural mismatch (3072 vs 4096 hidden,
+    5+20 vs 8+24 blocks), not a quality question.
+  • Mixing a "Flux.2 Klein 9B" LoRA and a "Flux.2 Klein 9B-base" LoRA in
+    one stack is fine — those strings describe the training checkpoint, not
+    incompatible tensor layouts. Set the sampler for the checkpoint
+    actually loaded.
+  • Run base-trained LoRAs on the distilled checkpoint when you want speed:
+    HF's klein LoRA guide reports the adapter loads there and typically
+    gives better results than on base, at 4 steps.
+  • A base-to-turbo / distillation-extraction LoRA is a whole-model delta,
+    not a content LoRA. Put it on the base checkpoint only, and treat its
+    strength as a steps/CFG dial rather than a style knob. Do not add it on
+    top of the already-distilled checkpoint.
+  • Read each LoRA's own published strength range instead of defaulting
+    everything to 1.0 — klein LoRAs in the wild ship very different ones
+    (0.5 start for the consistency LoRA, 0.8 for SOLRICKS' detail LoRA,
+    under 0.5 and sometimes 0.1 for the vafipas663 distillation delta).
+  • With a consistency or edit LoRA in the stack, excessive strength
+    suppresses the prompt: the consistency LoRA's article says that if
+    edits stop applying, lower it before changing anything else.
+  • Prefer runtime LoRA patching over baking a LoRA into the checkpoint:
+    ComfyUI's Model Merge / Model Save nodes have not detected Flux2 Klein
+    models (4B or 9B) since a March 2026 update — issue #13637, still open.
+  • The author never states which weight family this targets. Two different
+    questions, two different answers. Within 9B (distilled vs base — the
+    two CivitAI strings): not a real incompatibility. Both load through the
+    same Flux2KleinPipeline and have the same repo layout; for the 4B pair,
+    where both transformer configs are public, the distil
+
+───────────────────────────────────────────────────────────────────────────────
+  PROS
+───────────────────────────────────────────────────────────────────────────────
+  + Explicit trigger word(s) — hand-drawn sketch — so the effect can be
+    turned on and off from the prompt. [source: model metadata]
+  + Compact at 79 MB — cheap to stack with other LoRAs. [source: file size]
+  + Author published concrete recommended settings. [source: description]
+
+───────────────────────────────────────────────────────────────────────────────
+  CONS & CAVEATS
+───────────────────────────────────────────────────────────────────────────────
+  - Author never states which generation mode it was trained for; test
+    against your own workflow before relying on it. [source: description]
+  - Thin documentation — expect trial and error. [source: description
+    length]
+  - Modest adoption (361 downloads) — less community feedback to rely on.
+    [source: source stats]
+
+───────────────────────────────────────────────────────────────────────────────
+  COMPATIBILITY WARNING
+───────────────────────────────────────────────────────────────────────────────
+  Two different questions, two different answers. Within 9B (distilled vs
+  base — the two CivitAI strings): not a real incompatibility. Both load
+  through the same Flux2KleinPipeline and have the same repo layout; for the
+  4B pair, where both transformer configs are public, the distilled and base
+  configs are byte-for-byte identical in every shape field, and the 9B pair
+  is built the same way. HF's own klein LoRA guide states the workflow
+  directly: "Distilled models are step-compressed for fast inference; you
+  train against the base checkpoint and the adapter still loads on the
+  distilled model afterward", and "Applying the LoRA on the distilled model
+  typically gives better results than the base model, and it's faster" (that
+  post is written against 4B, but the family structure is the same). So the
+  two CivitAI strings record which checkpoint the author trained or previewed
+  against, not a hard compatibility wall. The real risk is a silent settings
+  mismatch: load a base-trained LoRA on the distilled checkpoint while
+  keeping the author's 20-50 step / CFG ~4 preview settings and the output
+  degrades badly (distilled wants 4 steps / guidance 1.0), or the reverse.
+  Nothing errors — it just looks wrong. Across sizes (9B vs 4B): a genuine
+  architectural mismatch, from the published transformer configs — different
+  hidden width (4096 vs 3072), different block counts (8+24 vs 5+20) and a
+  different text-conditioning dim (12288 vs 7680, following Qwen3-8B vs
+  Qwen3-4B). A 4B LoRA cannot patch a 9B model: expect missing keys and shape
+  errors, not degraded output. The exact ComfyUI error text is unverified. KV
+  variant: same 9B size and same 4-step distillation, but a distinct
+  checkpoint with its own pipeline; whether standard klein 9B LoRAs patch it
+  cleanly is not documented — see unverified.
+
+───────────────────────────────────────────────────────────────────────────────
+  AUTHOR'S DESCRIPTION
+───────────────────────────────────────────────────────────────────────────────
+  Klein 9B Works better with simple prompts.
+  Showcase images :
+  - Distilled model
+
+  - 768x1152
+
+  - Euler / normal / 4 steps / CFG 1
+
+  - Lora strength of 1-1.7
+
+───────────────────────────────────────────────────────────────────────────────
+  EXAMPLE MEDIA
+───────────────────────────────────────────────────────────────────────────────
+  6 preview(s) in ./example_images/ — community results for
+  this LoRA. Videos keep a matching _poster.jpg still frame.
+
+  No prompts are recorded below: CivitAI's public API flags these posts as
+  having generation metadata but does not return it, so the prompts cannot be
+  scraped. Open the model page to read them.
+
+  01_image.jpeg  (image, 768x1152)
+  02_image.jpeg  (image, 768x1152)
+  03_image.jpeg  (image, 768x1152)
+  04_image.jpeg  (image, 768x1152)
+  05_image.jpeg  (image, 768x1152)
+  06_image.jpeg  (image, 768x1152)
+
+───────────────────────────────────────────────────────────────────────────────
+  HOW TO USE (COMFYUI)
+───────────────────────────────────────────────────────────────────────────────
+  1. Diffusion model -> ComfyUI/models/diffusion_models/. Distilled:
+  flux-2-klein-9b-fp8.safetensors. Base: flux-2-klein-
+  base-9b-fp8.safetensors. (BF16 single-file weights
+  flux-2-klein-9b.safetensors / flux-2-klein-base-9b.safetensors from the
+  gated HF repos also work if you have ~29GB; community GGUF builds exist
+  from unsloth and leejet for lower-VRAM setups.)
+  2. Text encoder ->
+  ComfyUI/models/text_encoders/qwen_3_8b_fp8mixed.safetensors (from Comfy-
+  Org/vae-text-encorder-for-flux-klein-9b). This is the 9B family's encoder;
+  the 4B family uses qwen_3_4b.safetensors — do not mix them.
+  3. VAE -> ComfyUI/models/vae/flux2-vae.safetensors (shared across klein
+  sizes).
+  4. LoRA -> ComfyUI/models/loras/ as a plain .safetensors.
+  5. Load with Load Diffusion Model (UNETLoader) for the transformer,
+  CLIPLoader for the Qwen3 text encoder, VAELoader for the VAE.
+  docs.comfy.org ships prebuilt Flux.2 klein graphs under Workflow Templates
+  for both the distilled and base branches — start from those rather than
+  wiring by hand, since neither the docs nor BFL's repo state
+  sampler/scheduler/shift in prose; those values live in the template.
+  6. Insert LoraLoaderModelOnly between the diffusion-model loader and the
+  sampler. Model-only is correct for klein: the Qwen3 text encoder is not
+  trained by the standard klein trainers, so there is no text-encoder half to
+  patch.
+  7. Set steps and guidance to match the checkpoint you actually loaded, not
+  the one the LoRA was trained on: 4 steps / guidance 1.0 on distilled,
+  ~20-50 steps / guidance ~4.0 on base.
+  8. Note the 9B repos are gated on HuggingFace — you must accept the FLUX
+  Non-Commercial License on the repo page before any download, including
+  scripted ones.
+
+───────────────────────────────────────────────────────────────────────────────
+  Scraped by lorakit on 2026-09-05 03:55 UTC.
+  Stats and description are the author's; pros/cons are derived from the
+  evidence tagged beside each line.
+───────────────────────────────────────────────────────────────────────────────

hand-drawn-sketch/metadata.json

@@ -0,0 +1,140 @@
+{
+  "name": "Hand-drawn sketch",
+  "slug": "hand-drawn-sketch",
+  "base_model": "FLUX.2 [klein] 9B",
+  "media_type": "image",
+  "model_id": 2692445,
+  "version_id": 3023139,
+  "version_name": "Klein9B v1",
+  "source": "civitai",
+  "url": "https://civitai.com/models/2692445",
+  "repo_id": null,
+  "file_path": null,
+  "author": "Yofaraway",
+  "category": {
+    "key": "style",
+    "label": "Visual style"
+  },
+  "modalities": [],
+  "primary_modality": "unspecified",
+  "checkpoint_family": "unknown",
+  "trigger_words": [
+    "hand-drawn sketch"
+  ],
+  "recommended_settings": {
+    "Sampler steps": "4",
+    "Resolution": "768x1152",
+    "CFG / guidance": "1",
+    "Sampler": "Euler"
+  },
+  "chaining": {
+    "author": {
+      "stacks": "not stated",
+      "notes": "",
+      "avoid": [],
+      "requires": [],
+      "order": "",
+      "strength": "",
+      "stated": false
+    },
+    "general_rules": [
+      "Chain LoRAs by stringing LoraLoaderModelOnly nodes in series, model output into model input. There is no CLIP side to route — klein LoRAs are model-only because the Qwen3 text encoder is not trained (CivitAI's recipe fixes trainTextEncoder: false).",
+      "Only stack LoRAs of the same parameter size. 9B stacks with 9B; a 4B LoRA in a 9B stack is an architectural mismatch (3072 vs 4096 hidden, 5+20 vs 8+24 blocks), not a quality question.",
+      "Mixing a \"Flux.2 Klein 9B\" LoRA and a \"Flux.2 Klein 9B-base\" LoRA in one stack is fine — those strings describe the training checkpoint, not incompatible tensor layouts. Set the sampler for the checkpoint actually loaded.",
+      "Run base-trained LoRAs on the distilled checkpoint when you want speed: HF's klein LoRA guide reports the adapter loads there and typically gives better results than on base, at 4 steps.",
+      "A base-to-turbo / distillation-extraction LoRA is a whole-model delta, not a content LoRA. Put it on the base checkpoint only, and treat its strength as a steps/CFG dial rather than a style knob. Do not add it on top of the already-distilled checkpoint.",
+      "Read each LoRA's own published strength range instead of defaulting everything to 1.0 — klein LoRAs in the wild ship very different ones (0.5 start for the consistency LoRA, 0.8 for SOLRICKS' detail LoRA, under 0.5 and sometimes 0.1 for the vafipas663 distillation delta).",
+      "With a consistency or edit LoRA in the stack, excessive strength suppresses the prompt: the consistency LoRA's article says that if edits stop applying, lower it before changing anything else.",
+      "Prefer runtime LoRA patching over baking a LoRA into the checkpoint: ComfyUI's Model Merge / Model Save nodes have not detected Flux2 Klein models (4B or 9B) since a March 2026 update — issue #13637, still open.",
+      "The author never states which weight family this targets. Two different questions, two different answers. Within 9B (distilled vs base — the two CivitAI strings): not a real incompatibility. Both load through the same Flux2KleinPipeline and have the same repo layout; for the 4B pair, where both transformer configs are public, the distil"
+    ]
+  },
+  "pros": [
+    "Explicit trigger word(s) — hand-drawn sketch — so the effect can be turned on and off from the prompt. [source: model metadata]",
+    "Compact at 79 MB — cheap to stack with other LoRAs. [source: file size]",
+    "Author published concrete recommended settings. [source: description]"
+  ],
+  "cons": [
+    "Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description]",
+    "Thin documentation — expect trial and error. [source: description length]",
+    "Modest adoption (361 downloads) — less community feedback to rely on. [source: source stats]"
+  ],
+  "file": {
+    "name": "hand_drawn_sketch_klein9b.safetensors",
+    "size_bytes": 82879480,
+    "size_mb": 79.04,
+    "sha256": "96b1ce30cadd42bc15bb457117814dd0d31e19172e236f3038fc877129e869aa"
+  },
+  "stats": {
+    "downloads": 361,
+    "thumbs_up": 48,
+    "thumbs_down": 0
+  },
+  "popularity": {
+    "downloads": 361,
+    "thumbs_up": 48,
+    "thumbs_down": 0,
+    "comments": 1,
+    "approval": "100% positive (48/48)",
+    "age_days": 80.5,
+    "downloads_per_day": 4.48,
+    "rank_by_downloads": 192,
+    "rank_by_momentum": 127,
+    "of_total": 200,
+    "adoption_tier": "Niche",
+    "momentum": "Long tail"
+  },
+  "selection_score": 70.9,
+  "selection_reasons": [
+    "361 downloads",
+    "48 thumbs-up",
+    "top-ranked for its category"
+  ],
+  "example_media": [
+    {
+      "file": "01_image.jpeg",
+      "type": "image",
+      "resolution": "768x1152",
+      "prompt": ""
+    },
+    {
+      "file": "02_image.jpeg",
+      "type": "image",
+      "resolution": "768x1152",
+      "prompt": ""
+    },
+    {
+      "file": "03_image.jpeg",
+      "type": "image",
+      "resolution": "768x1152",
+      "prompt": ""
+    },
+    {
+      "file": "04_image.jpeg",
+      "type": "image",
+      "resolution": "768x1152",
+      "prompt": ""
+    },
+    {
+      "file": "05_image.jpeg",
+      "type": "image",
+      "resolution": "768x1152",
+      "prompt": ""
+    },
+    {
+      "file": "06_image.jpeg",
+      "type": "image",
+      "resolution": "768x1152",
+      "prompt": ""
+    }
+  ],
+  "tags": [
+    "drawing",
+    "style",
+    "strokes",
+    "sketch",
+    "pencil"
+  ],
+  "scraped_at": "2026-09-05T03:55:52+00:00",
+  "path": "flux2-klein-9b/hand-drawn-sketch"
+}