Anatomy Correction in Diffusion Models¶

Methods for detecting and fixing anatomy mutations (extra fingers, distorted hands, missing limbs) in diffusion-generated images. Covers post-processing fixes, during-sampling guidance, training-time solutions, and ComfyUI integration for FLUX Klein 9B.

Detection Tools¶

YOLO-based Hand/Face Detection (ComfyUI Impact Pack)¶

UltralyticsDetectorProvider node (from ComfyUI-Impact-Subpack)
Models: hand_yolov8s.pt, hand_yolov8n.pt (bbox), face_yolov8m.pt
Location: ComfyUI/models/ultralytics/bbox/

HADM - Human Artifact Detection¶

Dataset: HAD - 37,000+ images annotated for human artifact localization. Detects local (distorted faces/hands) and global (missing/extra limbs) artifacts. Generalizes across unseen generators.

MediaPipe Hand Landmarks¶

21 hand keypoints per hand. Finger tips at landmarks 4, 8, 12, 16, 20. Can validate finger count programmatically. Not a ComfyUI node natively - needs custom wrapper.

During-Sampling Solutions¶

NAG - Normalized Attention Guidance¶

The most promising during-sampling fix for FLUX/Klein. Training-free, works at inference time.

Package: ComfyUI-NAG (github.com/ChenDarYen/ComfyUI-NAG)
Supports: FLUX, Flux Kontext, Wan, HunyuanVideo, Chroma, SD3.5, SDXL

Key parameters for FLUX:

Tune nag_tau + nag_alpha first, then nag_scale. Acts as negative guidance improving structural coherence.

Flux2Klein-Enhancer (Klein 9B Only)¶

Repo: github.com/capitan01R/ComfyUI-Flux2Klein-Enhancer
Klein 9B ONLY - does not work with other FLUX variants

Modifies active text embedding region to improve prompt adherence. Makes anatomical prompt language "stick" better when Klein ignores it:

Combine with anatomy LoRA: enhancer makes the prompt effective, LoRA enforces the anatomy direction. NAG adds a third orthogonal axis - none cancel each other.

VSF - Value Sign Flip (arXiv 2508.10931)¶

Inference technique for negative guidance in few-step distilled models. Flips value signs in self-attention to achieve negative guidance effect without CFG. Works in models where standard negative prompts have no effect due to distillation. Complementary to NAG - both target different levels of the attention mechanism.

Other Guidance Methods¶

• PAG (Perturbed-Attention Guidance): fixes ~80% of finger errors in SDXL at scale=0.3. FLUX support NOT confirmed.
• SEG, SWG, PLADIS, TPG, FDG, MG: available via sd-perturbed-attention package, FLUX compatibility varies.

Post-Processing: Detect-and-Inpaint¶

HandFixer (ComfyUI Native)¶

One-click hand repair. Uses FLUX.1-Fill as inpainting backbone.

cd ComfyUI/custom_nodes
git clone https://github.com/Xiangyu-CAS/HandFixer
pip install -r HandFixer/requirements.txt
# Requires: ae.safetensors, clip_l.safetensors, t5xxl_fp16.safetensors, fluxl-fill-dev.safetensors

Supports realistic + anime, images from SDXL/DiT/Midjourney. FLUX-Fill recommended over FLUX-dev (better boundary handling).

HandRefiner (MeshGraphormer)¶

Most reliable approach for guaranteed correct finger count. Reconstructs 3D hand mesh -> depth map -> ControlNet-conditioned inpainting.

ComfyUI node: MeshGraphormer Hand Refiner (in comfyui_controlnet_aux)
Key params: detect_thr=0.6, presence_thr=0.6, control_strength=0.4-0.8

Control strength 1.0 causes texture loss. Use 0.4-0.8 range.

FaceDetailer (Impact Pack)¶

Key parameters:
  guide_size: 512-768 (reference size for enlarging detected regions)
  denoise: 0.3-0.5 (subtle fix) | 0.5-0.7 (moderate correction)
  noise_mask: True
  force_inpaint: True
  feather: 5-20px
  cycle: 1-2 (2 for badly corrupted faces)
  crop_factor: 3.0
  bbox_threshold: 0.5
  dilation: 0-20

For hands: use DetailerForEach with hand_yolov8s.pt, denoise 0.4-0.6 (lower than faces).

Universal Detailer¶

git clone https://github.com/ekakit/ComfyUI-UniversalDetailer

Extends FaceDetailer to hands and fingers via YOLO detection. Pipeline: YOLO detect -> SAM mask -> inpaint. Drop-in replacement when FaceDetailer is too face-focused.

Generic Inpainting Approach¶

1. Detect hands with hand_yolov8s.pt -> get mask
2. Expand mask with Mask Grow/Dilate node (10-20px)
3. Inpaint with FLUX Fill at denoise 0.75-0.90
4. Prompt override: "detailed hand, five fingers, correct anatomy, perfect hand"

Academic Methods (2024-2026)¶

Post-Processing Papers¶

Training-Time Papers¶

Evaluation / Detection Papers¶

Hand-Specific Foundation Model¶

FoundHand-10M: 10M hand images with 2D keypoints + segmentation masks. Capabilities: repose hands, transfer appearance, synthesize novel views, zero-shot fix malformed hands. License: CC BY-NC 4.0.

Hand Fix LoRAs¶

Concept Slider LoRA (Zero Images)¶

Text-based slider requiring no training data:

positive: "perfect anatomy, correct hands, five fingers"
unconditional: "broken hands, extra fingers, deformed anatomy"
# Training: 25-50 steps only, rank 4, alpha 1
# Total time: ~5 minutes

Pre-trained LoRAs for Klein 9B¶

Critical: klein_slider_anatomy uses a 1.0-4.0 weight scale, not 0-1. Weight 2.0 = minor fix, 3.0 = prominent fix, 4.0 = maximum. Stacking at 0.5-1.0 has no meaningful effect.

Stacking order for style + anatomy:

1. Style LoRA:              weight 0.7-1.0, strength_clip 0.4-0.6
2. klein_slider_anatomy v1.5: weight 2.0-3.0, strength_clip 0.3-0.5
3. DPO Klein 9B (optional): weight 0.4-0.6

CLIP weight trick: reduce strength_clip on anatomy LoRA more than strength_model.
anatomy LoRA model: 1.0-3.0, clip: 0.3-0.5
This lets anatomy correction affect denoising trajectory while limiting
its influence on text token processing (where style gets encoded).

Asian/Community Anatomy LoRAs¶

Available on Tensor.Art, Shakker.ai, LiblibAI for Klein 9B: - 手部肢体修复完善 V5 (Tensor.Art): body + limb repair, V5 with improved hand + torso - YesWenwen Hand Repair (Shakker): dedicated hand repair LoRA - Xian-T Hand Repair (LiblibAI): community-tested for hands/fingers

Training Data for Anatomy Correction¶

DPO Pipeline (Automated)¶

1. Take 5K-10K prompts with humans/hands
2. Generate 8 images per prompt with different seeds
3. Score with PickScore/HPSv2 for quality ranking
4. Best = winner, worst = loser
5. Optionally: use real photos as winners (hard stage)
6. Effort: ~40K-80K generations, ~2-3 hours on 4x GPU

Edit LoRA Pairs (Semi-Automated)¶

Method: HandRefiner as labeler

1. Generate batch of images with Klein 9B
2. Run HandRefiner/HandCraft (detect + fix)
3. Before = original, After = HandRefiner output
4. Auto-filter: keep only pairs where hand detector found issues
5. Need human QA pass on results

Summary Table¶

Klein 9B Best Anatomy Pipeline¶

Step 1 (before LoRAs): generation parameters

Recommended pipeline order (effectiveness ranking):

1. Generation: 8+ steps, simpler pose, res2s sampler
2. LoRA stack: klein_slider_anatomy v1.5 weight 2.0-3.0 + optional DPO at 0.5
3. During sampling: NAG (nag_sigma_end=0.75) + Flux2Klein-Enhancer (magnitude 1.15-1.25)
4. Post-gen detection: hand_yolov8s.pt + face_yolov8m.pt via Impact Pack
5. If hands bad: HandFixer or MeshGraphormer -> FLUX Fill inpaint at denoise 0.7
6. If face bad: FaceDetailer at denoise 0.4-0.5
7. Inpainting fallback: same Klein 9B + same full LoRA stack, mask only bad area, denoise 0.35-0.45
8. Quality gate: aesthetic score + hand detection confidence
9. If still bad: regenerate with new seed (up to 3-4 attempts)

Klein 9B Distilled vs Base for Anatomy¶

Installation¶

cd ComfyUI/custom_nodes

# Impact Pack (face/hand detection + detailing)
git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack
git clone https://github.com/ltdrdata/ComfyUI-Impact-Subpack

# HandFixer
git clone https://github.com/Xiangyu-CAS/HandFixer
pip install -r HandFixer/requirements.txt

# ControlNet Auxiliary (MeshGraphormer, OpenPose)
git clone https://github.com/Fannovel16/comfyui_controlnet_aux

# NAG (during-sampling guidance)
git clone https://github.com/ChenDarYen/ComfyUI-NAG

# Required models:
# ComfyUI/models/ultralytics/bbox/hand_yolov8s.pt
# ComfyUI/models/ultralytics/bbox/face_yolov8m.pt
# ComfyUI/models/unet/fluxl-fill-dev.safetensors

Gotchas¶

• NAG bug for FLUX: the ComfyUI-NAG package had a FLUX guidance degradation bug fixed in June 2025. Must use the updated version - older versions make FLUX outputs worse, not better.
• MeshGraphormer control_strength=1.0 kills texture: using full control strength from the depth map produces plasticky, over-smoothed hands. Use 0.4-0.8 range for realistic results.
• Klein distilled CFG must be 1.0: the distilled model has guidance baked in. CFG > 2.0 produces "deep-fried" artifacts. The base model supports CFG 3.5-5.0 for better anatomy control.
• 4B models have significantly worse anatomy than 9B: neither 4B variant (distilled or base) achieves acceptable anatomy quality for professional use. Budget for 9B.
• klein_slider_anatomy weight scale is 1-4, not 0-1: applying at weight 0.5-1.0 (typical LoRA range) has no meaningful effect. Use 2.0 for minor fixes, 3.0 for prominent artifacts.
• Inpainting fallback must use same model + same LoRA stack: using HandFixer (FLUX.1 Fill) on Klein-generated images creates style boundary mismatch. Use Klein 9B for the inpainting pass too.

See Also¶

• flux klein 9b inference - optimal inference settings including anatomy hierarchy
• face detection filtering pipeline - detection tools for quality gating
• face beautify edit lora - edit LoRA for facial correction
• diffusion lora training - training parameters for Klein LoRAs
• lora fine tuning for editing models - MMDiT editing model patterns