storyboard-sketch/info.txt

═══════════════════════════════════════════════════════════════════════════════
  STORYBOARD SKETCH
═══════════════════════════════════════════════════════════════════════════════

───────────────────────────────────────────────────────────────────────────────
  IDENTITY
───────────────────────────────────────────────────────────────────────────────
  Base model        FLUX.1 [dev]
  Generation mode   not declared by the author
  Checkpoint        not determined — verify before use
  Category          Visual style
  Author            bblink787
  Source            https://civitai.com/models/162118
  Provenance        CivitAI · model 162118 · version 869189

───────────────────────────────────────────────────────────────────────────────
  WEIGHTS
───────────────────────────────────────────────────────────────────────────────
  File              Storyboard_sketch--FLUX.safetensors
  Version           FLUX
  Size              292.3 MB
  Format            SafeTensor / unknown precision
  SHA256            e35849ea0b7cb822f1fbc5a2cb6eaee6348bc95e4ee9e9e4dc175812c1f2e2b8
  Published         2024-09-19

───────────────────────────────────────────────────────────────────────────────
  TRIGGER WORDS
───────────────────────────────────────────────────────────────────────────────
  › "Storyboard sketch"

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

  LoRA strength       0.8
  Sampler steps       30
  CFG / guidance      5
  Sampler             Euler_ancestral

───────────────────────────────────────────────────────────────────────────────
  ADOPTION & POPULARITY
───────────────────────────────────────────────────────────────────────────────
  Downloads               11,081
  Rating                  100% positive (982/982)
  Comments                7
  Published               715 days ago
  Download rate           4.4/day since release

  Adoption rank           #135 of 150 in this collection  (Niche)
  Momentum rank           #95 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 FLUX.1 [dev] ones, not anything specific to this LoRA.

  General FLUX.1 [dev] rules that apply here (see REFERENCE.txt):

  • Chain by wiring MODEL (and CLIP, if using LoraLoader) from one loader
    into the next; ComfyUI's official multi-LoRA tutorial shows sequential
    Load LoRA nodes with per-node strength_model.
  • Practically, what changes results is the strengths, not the node order
    — but see unverified for the strict-commutativity claim.
  • Put an accelerator LoRA first and reduce content-LoRA strengths under
    it. This is community convention, not documented behaviour.
  • Do not stack two accelerator LoRAs (Hyper-SD + Turbo-Alpha). Neither
    publisher documents it and their step schedules and trained guidance
    points differ.
  • The author never states which weight family this targets. Almost always
    SILENT DEGRADATION, not a hard error. Two verified mechanisms in
    ComfyUI: (1) Key miss — an unmatched LoRA key is skipped with
    `logging.warning("lora key not loaded: {}")` at comfy/lora.py:93 and
    generation proceeds with that layer unpatched. (2) Shape mismatch — co

───────────────────────────────────────────────────────────────────────────────
  PROS
───────────────────────────────────────────────────────────────────────────────
  + Widely adopted — 11,081 downloads. [source: CivitAI stats]
  + Positively received — 982 thumbs-up with no down-votes. [source:
    CivitAI stats]
  + Explicit trigger word(s) — Storyboard sketch — so the effect can be
    turned on and off from the prompt. [source: model metadata]
  + Compact at 292 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]

───────────────────────────────────────────────────────────────────────────────
  COMPATIBILITY WARNING
───────────────────────────────────────────────────────────────────────────────
  Almost always SILENT DEGRADATION, not a hard error. Two verified mechanisms
  in ComfyUI: (1) Key miss — an unmatched LoRA key is skipped with
  `logging.warning("lora key not loaded: {}")` at comfy/lora.py:93 and
  generation proceeds with that layer unpatched. (2) Shape mismatch —
  comfy/weight_adapter/lora.py wraps the
  `torch.mm(...).reshape(weight.shape)` merge in try/except and on failure
  logs `ERROR <name> <key> <exception>`, leaving that weight untouched and
  continuing. So a mismatched LoRA yields a normal-looking image with the
  LoRA partly or wholly absent; you only find out by reading the terminal.
  ComfyUI also has explicit `pad_tensor_to_shape` handling
  (comfy/weight_adapter/base.py) so BFL's own Canny/Depth dev LoRAs can widen
  `img_in` when applied across differing input-channel counts — that path is
  deliberate. The genuinely dangerous case is flux1-kontext-dev and
  flux1-krea-dev: shapes match exactly, so nothing is logged and a wrong-
  family LoRA merges cleanly while producing subtly wrong output. Hard errors
  are essentially limited to loading a non-Flux LoRA (SDXL/SD1.5) whose key
  namespace has no overlap, and to bnb-NF4 checkpoints where the quantized
  weight cannot be patched without conversion.

───────────────────────────────────────────────────────────────────────────────
  AUTHOR'S DESCRIPTION
───────────────────────────────────────────────────────────────────────────────
  Z-Image Turbo Version: Note the trigger words on this version. Z-Image
  Turbo is very sensitive to the word "storyboard" and will generate multiple
  frames in the same image, so that had to be eliminated. The words "digital
  sketch" are important to override Z-Image's strong tendency to generate
  photographic content.

  Using the sampler/scheduler: Euler_ancestral/
  FlowMatchEulerDiscreteScheduler will give results that are closest to the
  training images.

  Start at a strength of 0.8 or 0.85 to get a strong effect with less anatomy
  problems.

  Illustrious Version: This version was trained on the same updated dataset
  used for the Flux and Pony versions. Example images were created with the
  Hassaku XL (Illustrious) checkpoint (v1.3 - Style A) Euler a/Simple, CFG 5,
  30 steps. Many, many thanks to @ChronoKnight for his encouragement, wisdom,
  and buzz! This wouldn't have happened without him. This is my first attempt
  at an Illustrious model so feedback is, of course, welcome.

  FLUX Version: Start out at a strength of 0.8 . If you go up to 1, be
  prepared for some funky mutations. Specify motion blur , if you want it,
  and choose your camera view ( closeup, over the shoulder shot, aerial shot,
  medium shot, etc. )
  I ran this training 7 different times/ways before I was happy(ish) with the
  results. Flux is a wild animal and really fought this style for some
  reason. All training images were black and white, yet this produces many
  color images. Occasionally it will produce photo-real images (stronger
  epochs only introduced more mutations and hallucinations, not fewer photo-
  reals). So, it is what it is ;)

  SDXL Version: This LoRA was trained on SDXL Base using 60 grayscale
  storyboard sketches and character portraits at 21:9, 16:9, and 1:1 aspect
  ratios. At full strength you get the most abstract sketches - good for
  leaving the fine details of a scene to the imagination. At around 0.8
  strength you get much more coherence and prompt faithfulness. You can go
  lower from there to get more detailed and realistic sketches.

  Strength of 1.0 : Strongest styling, less coherence and prompt faithfulness
  Strength of 0.8 : Good styling, medium coherence and prompt faithfulness
  Strength of 0.5 : Conventional styling, best coherence and prompt
  faithfulness

  Pony Version: This version was trained on an updated dataset and I'm really
  happy with how it turned out. Example images are all at a CFG of 10 or 12
  to get more contrast, and Euler/Normal or Euler a/karras. Same strength
  guidelines as the SDXL version.

───────────────────────────────────────────────────────────────────────────────
  VERSION NOTES
───────────────────────────────────────────────────────────────────────────────
  Start out at a strength of 0.8. If you go up to 1, be prepared for some
  funky mutations.
  Specify motion blur, if you want it, and choose your camera view (close-up,
  over the shoulder shot, from below, etc.)

───────────────────────────────────────────────────────────────────────────────
  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, 832x1216)
  03_image.jpeg  (image, 832x1216)
  04_image.jpeg  (image, 832x1216)
  05_image.jpeg  (image, 832x1216)
  06_image.jpeg  (image, 832x1216)

───────────────────────────────────────────────────────────────────────────────
  HOW TO USE (COMFYUI)
───────────────────────────────────────────────────────────────────────────────
  1. Install the base model. Either (a) split files — `flux1-dev.safetensors`
  (23.8 GB bf16, diffusion model ONLY) in `ComfyUI/models/diffusion_models/`,
  `clip_l.safetensors` (250 MB) and `t5xxl_fp16.safetensors` (9.79 GB, or
  `t5xxl_fp8_e4m3fn_scaled.safetensors` 5.16 GB for low VRAM) in
  `ComfyUI/models/text_encoders/`, and `ae.safetensors` (335 MB) in
  `ComfyUI/models/vae/`; or (b) the all-in-one — `flux1-dev-fp8.safetensors`
  (17.25 GB, bundles text encoders + VAE) in `ComfyUI/models/checkpoints/`,
  loaded with CheckpointLoaderSimple. These paths are the ones in ComfyUI's
  official Flux.1 dev tutorial. black-forest-labs/FLUX.1-dev is auto-gated (a
  raw fetch of its LICENSE.md returns "Access to model ... is restricted");
  the ungated Comfy-Org/flux1-dev and comfyanonymous/flux_text_encoders
  repackages are what ComfyUI's templates link to.

  2. Drop the LoRA `.safetensors` into `ComfyUI/models/loras/`. Subfolders
  are allowed and show as `subdir/name.safetensors`. No config file, no
  restart — use the node's refresh.

  3. Wire the base graph, matching the shipped template: UNETLoader
  (weight_dtype `default`, or `fp8_e4m3fn` to halve VRAM) -> DualCLIPLoader
  with type `flux` (clip_l + t5xxl) -> CLIPTextEncode -> FluxGuidance (3.5)
  -> KSampler (steps 20, cfg 1.0, euler, simple) with EmptySD3LatentImage
  1024x1024 -> VAEDecode -> SaveImage. Negative conditioning is a
  ConditioningZeroOut off the same CLIPTextEncode. (The shipped template
  omits FluxGuidance entirely and relies on model_base.py's 3.5 fallback;
  adding the node is how you change guidance.)

  4. Insert the LoRA node between the model loader and the sampler. Use
  **LoraLoader** ("Load LoRA", model + clip in/out) when the LoRA contains
  text-encoder weights — ComfyUI's comfy/lora.py maps `lora_te1_*` (CLIP-L)
  and `lora_te2_*` keys. Use **LoraLoaderModelOnly** (model in/out only) for
  model-only LoRAs; this is what ComfyUI's official Flux ControlNet tutorial
  uses for `flux1-depth-dev-lora.safetensors`. Loading a model+clip LoRA with
  LoraLoaderModelOnly drops the text-encoder half.

  5. Quantized paths. GGUF: install the ComfyUI-GGUF custom node, put
  `flux1-dev-Q*.gguf` in `ComfyUI/models/unet/` (the README's stated path),
  load with the GGUF Unet loader; the README says "LoRA loading is
  experimental but it should work with just the built-in LoRA loader
  node(s)". Nunchaku SVDQuant INT4: install ComfyUI-nunchaku and use its
  Nunchaku FLUX LoRA Loader.

  6. Verify from the terminal, not the image. A LoRA that failed to apply
  prints `lora key not loaded: <key>` or `ERROR <name> <key> <exception>` and
  then generates a perfectly normal-looking image without it.

───────────────────────────────────────────────────────────────────────────────
  Scraped by lorakit on 2026-09-05 02:06 UTC.
  Stats and description are the author's; pros/cons are derived from the
  evidence tagged beside each line.
───────────────────────────────────────────────────────────────────────────────