tentacle-pit/info.txt

═══════════════════════════════════════════════════════════════════════════════
  TENTACLE PIT
═══════════════════════════════════════════════════════════════════════════════

───────────────────────────────────────────────────────────────────────────────
  IDENTITY
───────────────────────────────────────────────────────────────────────────────
  Base model        Krea 2
  Generation mode   Instruction / reference editing (community, not official), Text-to-image
  Checkpoint        Krea 2 RAW (krea/Krea-2-Raw)
  Category          Detail / anatomy fixer
  Author            mrkakapopoloch
  Source            https://civitai.com/models/445957
  Provenance        CivitAI · model 445957 · version 3145500

───────────────────────────────────────────────────────────────────────────────
  WEIGHTS
───────────────────────────────────────────────────────────────────────────────
  File              krea2_z_captions_62b8e17c_lora_3000_v1.safetensors
  Version           v1.0 K2
  Size              109.0 MB
  Format            SafeTensor / unknown precision
  SHA256            07201197e8e68b16018e4c04208f5dbc17e47149101bef19314dc15b7f47cbb3
  Published         2026-07-19

───────────────────────────────────────────────────────────────────────────────
  TRIGGER WORDS
───────────────────────────────────────────────────────────────────────────────
  › "t3nt4cl3"

───────────────────────────────────────────────────────────────────────────────
  RECOMMENDED SETTINGS
───────────────────────────────────────────────────────────────────────────────
  Taken verbatim from the author's description:

  LoRA strength       0.5–0.8

───────────────────────────────────────────────────────────────────────────────
  ADOPTION & POPULARITY
───────────────────────────────────────────────────────────────────────────────
  Downloads               9,338
  Rating                  100% positive (674/675)
  Comments                0
  Published               47 days ago
  Download rate           19.5/day since release

  Adoption rank           #59 of 150 in this collection  (Moderate)
  Momentum rank           #99 of 150 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 Krea 2 ones, not anything specific to this LoRA.

  General Krea 2 rules that apply here (see REFERENCE.txt):

  • The most important rule is not about chaining but about which
    checkpoint you chain onto: the official README says in bold "TRAIN on
    Raw and RUN on Turbo", and "We highly recommend using RAW for training
    LoRAs and applying them on Turbo for inference." LoRAs trained on Raw
    are stated to work well on Turbo.
  • Chain with LoraLoaderModelOnly nodes in series (model -> model). Krea 2
    LoRAs are DiT-only, so there is no CLIP side to chain and the text-
    encoder branch stays untouched however many LoRAs you add.
  • Krea 2 LoRAs stack in practice and the community treats multi-LoRA as
    routine - style + character, style + slider, realism + character.
    Several CivitAI Krea 2 authors state their LoRAs stack cleanly; one
    writes "I recently stacked over 20 of my own LoRAs on a single
    generation and the output still came out great."
  • Reduce per-LoRA strength when stacking. CivitAI Krea 2 authors converge
    on dropping from ~1.0 to roughly 0.6-0.9 once more than one adapter is
    loaded. The CivitAI 12GB training guide also notes rank affects
    composability: "rank 8 produced a lighter style influence that was
    easier to stack with other LoRAs, rank 16 provided a useful balance
    between concept strength and composability, and rank 32 was better
    suited to a precise object or concept intended to take priority over
    other stacked LoRAs."
  • Accelerator LoRAs are position-sensitive in intent, not in node order:
    a Raw-to-Turbo delta adapter belongs on the Raw checkpoint, a
    sub-8-step LoRA belongs on Turbo (or on Raw already carrying a Raw-to-
    Turbo adapter). Chaining Raw -> Raw-to-Turbo adapter -> 4-step LoRA is
    explicitly recommended by the 4-step LoRA's author.
  • Whatever you stack, the schedule must stay Turbo's: cfg 1.0 in ComfyUI
    (guidance 0.0 in diffusers) and mu/shift 1.15. The 4-step LoRA's card
    says "Keep mu = 1.15. The training targets are anchored to that grid; a
    different shift moves the [sampler off them]"; the same reasoning
    applies to any Turbo-trained adapter in the chain.
  • Reference/edit LoRAs are designed to be stacked under ordinary subject
    and style LoRAs - the Identity Edit card states "Composes with your
    LoRAs: character/body/style LoRAs stack on top and steer the prior" and
    recommends a stacked subject LoRA when unusual subjects drift toward
    the base prior.
  • Krea 2 LoRAs cannot be chained with LoRAs for any other base model.
    Many CivitAI pages carry multi-model titles like "(Krea-2 + ZIT)" or
    "[Flux | ZIB | Krea2]" - those are separate files for separate
    architectures shipped under one model page, not a combined adapter.
  • To combine two full Krea 2 checkpoints rather than adapters, ComfyUI
    ships a dedicated ModelMergeKrea2 node (added in v0.27.0) with per-
    section weights for first., tmlp., txtmlp., tproj.,
    txtfusion.projector, txtfusion.layerwise_blocks.0-1,
    txtfusion.refiner_blocks.0-1, blocks.0-27 and last.
  • Undistilled base checkpoint, no step or CFG distillation and no post-
    training. The card says "its not recommended for inference use. Its a
    good base for finetuning or post-training". Official settings: 52
    steps, CFG 3.5, trained up to 1K. This is where Krea tells you to train
    LoRAs.

───────────────────────────────────────────────────────────────────────────────
  PROS
───────────────────────────────────────────────────────────────────────────────
  + Widely adopted — 9,338 downloads. [source: CivitAI stats]
  + Positively received — 674 thumbs-up, 1 down. [source: CivitAI stats]
  + Explicit trigger word(s) — t3nt4cl3 — so the effect can be turned on
    and off from the prompt. [source: model metadata]
  + Author documents use across several modes (EDIT, T2I). [source:
    description]
  + Compact at 109 MB — cheap to stack with other LoRAs. [source: file
    size]
  + Author published concrete recommended settings. [source: description]

───────────────────────────────────────────────────────────────────────────────
  CONS & CAVEATS
───────────────────────────────────────────────────────────────────────────────
  (none recorded)

───────────────────────────────────────────────────────────────────────────────
  COMPATIBILITY WARNING
───────────────────────────────────────────────────────────────────────────────
  Silent degradation, never a hard error, between Raw and Turbo. The two
  share one architecture - ai-toolkit's config is commented "The reference
  'single_mmdit_large_wide' architecture (oss_raw / oss_turbo share it)" - so
  every LoRA tensor shape matches on both and ComfyUI/diffusers will load an
  adapter against either without complaint. The lvladikov 4-step LoRA card
  states this for the reverse direction: "Every layer it targets also exists
  in Krea 2 Raw, so it will load there without complaint - but that is a side
  effect of the shared architecture, not a supported mode", and reports
  mixed, subject-dependent results with malformed subjects (duplicated heads,
  fused limbs) at low step counts, needing 14+ steps with Raw's CFG on before
  output is coherent. The intended direction (train on Raw, run on Turbo) is
  officially blessed. The more common practical failure is a settings
  mismatch rather than a weight mismatch: running a Turbo-targeted LoRA at
  Raw's 52 steps + CFG 3.5, or leaving CFG enabled on Turbo, degrades output
  with no warning. Loading a Krea 2 LoRA against a different model family
  entirely (Flux, Qwen-Image, Z-Image) is also not an error - ComfyUI logs
  "lora key not loaded: {key}" per key and generates as if no LoRA were
  applied.

───────────────────────────────────────────────────────────────────────────────
  AUTHOR'S DESCRIPTION
───────────────────────────────────────────────────────────────────────────────
  Due to filters im not able to upload proper Preview images for the models,
  will try to find a solution.
  F2K 9b: Works best with strength 1-1.5. Works for T2I but also works great
  for I2I/Edit F2K 4b: Not really recommend to use, as it has extreme body
  proportions issues, but 1 out of 5 images is decent. Lora training on
  civitai broke down at 9epochs. Will retrain soon. ZiT: Trained on 350
  images, mainly anime. Has some anatomy problems sometimes, but with
  strength 0.5-0.8 can put out some great images.

  Flux1Dev and below:
  Trained on 940 images containing:
  Tentacles,Slime,Insects,Assimilation,Pregnancy,Lacatation,oviposition
  works good from 0,3-1
  tested on Pony as also everclear and other merges
  SDXL version works fine on most base models.

───────────────────────────────────────────────────────────────────────────────
  EXAMPLE MEDIA
───────────────────────────────────────────────────────────────────────────────
  4 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, 1088x1920)
  02_image.jpeg  (image, 2048x1792)
  03_image.jpeg  (image, 1472x2048)
  04_image.jpeg  (image, 3840x2176)

───────────────────────────────────────────────────────────────────────────────
  HOW TO USE (COMFYUI)
───────────────────────────────────────────────────────────────────────────────
  1. Update ComfyUI to at least v0.26.0 (released 2026-06-23), which added
  local Krea2 support. For reference-image / style-reference / edit LoRAs you
  need at least v0.29.0 (2026-07-29), which added regular and timestep-zero
  reference images for Krea 2. ModelMergeKrea2 arrived in v0.27.0
  (2026-06-30).

  2. Put the base stack in place (files from huggingface.co/Comfy-Org/Krea-2,
  which is not gated):
     - ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors
  (documented recommended default; the repo also holds krea2_turbo_bf16,
  _int8_convrot, _mxfp8, _nvfp4 and krea2_raw_bf16 / _fp8_scaled /
  _int8_convrot)
     - ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (bf16
  also available)
     - ComfyUI/models/vae/qwen_image_vae.safetensors

  3. Put the LoRA .safetensors in ComfyUI/models/loras/ - flat, no per-model
  subfolder required.

  4. Build or load the graph. The official image_krea2_turbo_t2i.json
  template is: UNETLoader (krea2_turbo_fp8_scaled.safetensors) ->
  LoraLoaderModelOnly -> KSampler; CLIPLoader
  (qwen3vl_4b_fp8_scaled.safetensors, type "krea2") -> text encode ->
  KSampler positive, with ConditioningZeroOut into negative; VAELoader
  (qwen_image_vae.safetensors) -> VAEDecode -> SaveImage; EmptyLatentImage
  1024x1024 driven by ResolutionSelector. An int8 sibling template
  (image_krea2_turbo_t2i_int8.json) also ships.

  5. Load the LoRA with LoraLoaderModelOnly, not LoraLoader - both official
  templates do. Krea 2 LoRAs patch the DiT only; note that the DiT's own
  text_fusion / txtfusion blocks are a normal LoRA target (the official style
  LoRAs patch them), but the Qwen3-VL text encoder itself is not. Set
  strength 1.0 for Krea's official style LoRAs (the shipped template uses
  0.8) and add the LoRA's trigger word to the prompt.

  6. Sampler settings: steps 8, cfg 1, sampler euler, scheduler simple,
  denoise 1.0. Do not add a ModelSampling node for plain t2i - ComfyUI
  already applies shift 1.15 from Krea2.sampling_settings.

  7. For style reference, use the dedicated template: UNETLoader
  krea2_turbo_int8_convrot.safetensors, LoraLoaderModelOnly
  krea2_style_reference.safetensors at 1.0, ModelSamplingFlux (1.15, 0.5),
  TextEncodeQwenImageEditPlus fed by LoadImage,
  FluxKontextMultiReferenceLatentMethod set to index_timestep_zero, and
  SamplerCustomAdvanced with KSamplerSelect euler + BasicScheduler simple/8 +
  CFGGuider cfg 1.

  8. For instruction editing (Krea 2 Identity Edit), stock nodes are not
  enough: install the ComfyUI-Krea2Edit node pack, which ships its own
  workflows, and run Turbo at 8-12 steps / cfg 1 (or Raw at 20 steps / cfg 3
  for removals) with LoRA strength 1.0.

  9. Low-VRAM alternative: install city96's ComfyUI-GGUF and use a community
  GGUF build (realrebelai/KREA-2_GGUFs, quantized from Turbo;
  molbal/krea2-gguf, quantized from Raw) with the Unet Loader (GGUF) node in
  place of UNETLoader. There is no official GGUF release. LoRA loading is
  unchanged.

───────────────────────────────────────────────────────────────────────────────
  Scraped by lorakit on 2026-09-05 00:14 UTC.
  Stats and description are the author's; pros/cons are derived from the
  evidence tagged beside each line.
───────────────────────────────────────────────────────────────────────────────