Skip to content

face-fixer.md

Latest commit

 

History

History
154 lines (115 loc) · 5.81 KB

face-fixer.md

File metadata and controls

154 lines (115 loc) · 5.81 KB
Raw

Face Fixer (MEC)

End-to-end face detail node. YOLO11 face detection → per-face crop → optional AI upscale → per-face KSampler → smart blend back. Behaves like the legacy "Forbidden Vision Fixer" but with Impact-Pack wildcard syntax and ComfyUI-native sampling.

Display name: Face Fixer (MEC) Class: MECFaceFixer Category: MEC/Paint

What it does

Iterates over every detected face in every frame and runs an isolated KSampler pass at higher resolution, then blends the result back over the original with feather + color match + lightness rescue. Per-face prompts are supported via wildcard tokens.

Works on single images and video batches identically — each frame is processed independently, but the seed seeds advance per face so the output is reproducible.

Required parameters

Parameter Default Range / Choices Description
image — IMAGE (B,H,W,C) Source frame(s)
model — MODEL Diffusion model used for the per-face KSampler
positive / negative — CONDITIONING Base conditioning. Per-face wildcard prompts override
vae — VAE VAE for encode/decode of crops
face_model none YOLO11 .pt / .onnx from models/ultralytics/bbox/ Face detector. Set none to use the optional mask input
confidence 0.5 0.05–0.95 Detection threshold
max_faces 8 0–32 (0 = all) Cap per frame
crop_padding 1.4 1.0–3.0 Bbox padding multiplier so the sampler sees context
crop_resolution 768 256–2048 (step 64) Resize each crop's longer side before sampling
denoise 0.4 0.0–1.0 Per-face denoise strength (0.3 subtle, 0.7 aggressive reshape)
steps 20 1–100 Sampling steps per face
cfg 6.0 0–30 CFG scale
sampler_name euler KSampler list Sampler algorithm
scheduler normal KSampler list Sigma schedule
seed 0 int Base seed; each face gets seed + face_index
blend_softness 6.0 0–64 px Feather radius on per-face blend mask
mask_dilate 4 -32–32 Dilate (>0) / erode (<0) of blend mask
color_match true bool Reinhard mean+std colour match per face
lightness_rescue true bool Lift CIE LAB L if the sample comes back darker than the original
differential_diffusion true bool Weight the blend by abs(orig − sampled) so unchanged pixels stay sharp

Optional parameters

Parameter Description
mask Manual face mask. Used when face_model='none' or detection is empty
upscale_model UPSCALE_MODEL applied to faces below crop_resolution before sampling
face_positive_prompt Per-face positive prompt (wildcard syntax below). Empty = use base positive
face_negative_prompt Same syntax for negatives

Wildcard syntax (face_positive_prompt / face_negative_prompt)

Token Effect
[SEP] Separates per-face prompts. Order = detection order
[ASC] Order faces left-to-right
[DSC] Order faces right-to-left
[ASC-SIZE] Order faces small-to-large
[DSC-SIZE] Order faces large-to-small
[SKIP] Leave that face untouched (no sampling)

Examples

red lipstick [SEP] blue eyes [SEP] [SKIP]

Three faces: face 1 gets "red lipstick", face 2 "blue eyes", face 3 skipped.

[DSC-SIZE] hero glamour shot [SEP] background extra [SEP] background extra

Largest face = "hero glamour shot", smaller faces = "background extra".

Outputs

Output Type Description
image IMAGE Frame(s) with detailed faces blended back
face_mask MASK Combined mask of every processed face
info_json STRING Per-face metadata (bbox, score, prompt, denoise)

Use cases

Image generation — auto face hi-fix

KSampler ──▶ VAE Decode ──▶ MECFaceFixer (denoise=0.35, crop_resolution=1024) ──▶ Save

Run a normal txt2img then auto-detail the face at higher resolution without touching the rest of the image.

Video generation — per-frame face polish

Wan2.2 Animate ──▶ MECFaceFixer (face_model=face_yolo11n, denoise=0.3) ──▶ VHS Combine

Each frame is independently detailed. Use a low denoise (0.25–0.35) to preserve identity and reduce flicker.

Multi-character scenes

… ──▶ MECFaceFixer
       face_positive_prompt = "[DSC-SIZE] cinematic glamour [SEP] sharp eyes [SEP] [SKIP]"
       max_faces = 3

Largest face gets glamour treatment, second face just eye sharpening, third face skipped (e.g. background blur).

Recipes

Scenario Settings
Subtle hi-fix denoise=0.25, steps=20, crop_resolution=768, differential_diffusion=true
Aggressive reshape denoise=0.65, steps=30, cfg=7.5, crop_padding=1.6
Tiny faces in wide shot crop_padding=1.8, upscale_model=4xUltraSharp, crop_resolution=1024
Anti-flicker on video denoise≤0.35, seed=fixed, differential_diffusion=true, lightness_rescue=true
Identity-preserving polish color_match=true, lightness_rescue=true, differential_diffusion=true

Troubleshooting

Symptom Cause Fix
face_model dropdown empty No YOLO11 weights installed Place .pt in models/ultralytics/bbox/ (e.g. face_yolov8n.pt works too)
No faces detected Threshold too high / faces too small Lower confidence to 0.3, raise crop_padding to 1.8
Faces look "washed out" Color match too aggressive on small faces Disable color_match, keep lightness_rescue
Flicker on video High denoise per frame Drop to 0.25–0.35; enable differential_diffusion
Identity drift Denoise too high or wildcard prompt too strong Lower denoise, simplify wildcard