replace-subject-klein-4b-9b/info.txt

═══════════════════════════════════════════════════════════════════════════════
  REPLACE SUBJECT [KLEIN 4B & 9B]
═══════════════════════════════════════════════════════════════════════════════

───────────────────────────────────────────────────────────────────────────────
  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            ReltivlyObjectv
  Source            https://civitai.com/models/2508392
  Provenance        CivitAI · model 2508392 · version 2819585

───────────────────────────────────────────────────────────────────────────────
  WEIGHTS
───────────────────────────────────────────────────────────────────────────────
  File              KleinBase9B_ReplaceSubject.safetensors
  Version           Flux.2 Klein 9B (v1.0)
  Size              79.0 MB
  Format            SafeTensor / unknown precision
  SHA256            0fd02736097f1570da1ea0c0d8e5e584e77e2b0202eb7a8880629d3373850080
  Published         2026-03-31

───────────────────────────────────────────────────────────────────────────────
  TRIGGER WORDS
───────────────────────────────────────────────────────────────────────────────
  › "replace the person in Image 1 with the person in Image 2"
  › "replace the man in Image 1 with the man in Image 2"
  › "replace the woman in Image 1 with the woman in Image 2"

───────────────────────────────────────────────────────────────────────────────
  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               2,457
  Rating                  100% positive (151/151)
  Comments                0
  Published               157 days ago
  Download rate           13.3/day since release

  Adoption rank           #90 of 150 in this collection  (Modest)
  Momentum rank           #42 of 150 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 — 2,457 downloads. [source: CivitAI stats]
  + Positively received — 151 thumbs-up with no down-votes. [source:
    CivitAI stats]
  + Explicit trigger word(s) — replace the person in Image 1 with the
    person in Image 2, replace the man in Image 1 with the man in Image 2,
    replace the woman in Image 1 with the woman in Image 2 — 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]

───────────────────────────────────────────────────────────────────────────────
  CONS & CAVEATS
───────────────────────────────────────────────────────────────────────────────
  - Author labels this experimental — results vary shot to shot. [source:
    description]
  - 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
───────────────────────────────────────────────────────────────────────────────
  Allows you to replace a subject in the first control image with a subject
  in the second control image.
  The model will skew towards preserving existing details if the secondary
  control image does not include the relevant details and the prompt leaves
  them unspecified. For example, if you are using a head/torso image as the
  new subject, but the original image was zoomed out more and the person
  being replaced has jeans, the new person will also be wearing jeans unless
  you specify otherwise.
  For best results, try to have comparable framing, camera angle, and number
  of subjects in both images.
  replace the person in Image 1 with the person in Image 2
  replace the man in Image 1 with the man in Image 2
  replace the woman in Image 1 with the woman in Image 2 Model Comparisons:
  - Copy Pose will preserve the subject and background in Control 1, while
  bringing the framing/pose of Control 2

  - Replace Subject will preserve the framing/pose and background in Control
  1, while bringing the subject of Control 2

  - Replace Scene will preserve the subject and pose in Control 1, while
  bringing the background scene of Control 2

  California AB 2013 Training Data Disclosure
  - This LoRA was fine-tuned using visual data consisting entirely of
  synthetic still images. The training data may include copyrighted material
  owned by third parties. No training data was licensed or purchased. This
  LoRA is provided for non-commercial use only under the terms of its
  distribution.

  - The dataset consists of over 100 image sets (over 300 images total). This
  is the total used for training after low quality sets were culled. Dataset
  was created in 2026.

  - Image data was processed through standard resizing, cropping,
  normalization, and labeling steps. Synthetic images were included as part
  of the training dataset.

  - This model is intended for non-commercial, experimental, and educational
  use. Generated outputs may reflect copyrighted visual styles or themes
  associated with the underlying training data. Users are responsible for
  ensuring compliance with applicable copyright law, other intellectual
  property laws, and all other applicable laws.

───────────────────────────────────────────────────────────────────────────────
  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, 1024x640)
  02_image.jpeg  (image, 2148x3172)
  03_image.jpeg  (image, 832x1264)
  04_image.jpeg  (image, 1049x1988)
  05_image.jpeg  (image, 1852x2876)
  06_image.jpeg  (image, 928x1120)

───────────────────────────────────────────────────────────────────────────────
  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 03:25 UTC.
  Stats and description are the author's; pros/cons are derived from the
  evidence tagged beside each line.
───────────────────────────────────────────────────────────────────────────────