Ultimate Face Fix

ComfyUI's face-fix story usually ends in one of two places: a fixed restoration model (GFPGAN, CodeFormer) that smooths faces into plastic, or Impact Pack's FaceDetailer, which works but makes you wrestle a two-dozen-input node. Ultimate Face Fix is the new kid that splits the difference. It does what ADetailer did for the A1111 crowd - detect faces, give each one its own img2img pass, blend it back - but the "fix" is done by your checkpoint, VAE, prompt, and sampler, not by a generic face model. The name is a promise, and it mostly keeps it.

The mechanism is a detect → crop → repair → mask → blend pipeline, all inside one node:

1. A YOLO face detector finds every face (or just the largest, or the first N).
2. Each face is cropped square with context around it (context_scale, default 1.8) and resized to target_resolution (default 1024).
3. Each crop is re-rendered as img2img using the model, vae, positive, negative, sampler, and scheduler you connected, at whatever denoise you set.
4. The SegFace parser builds a precise semantic mask - which face parts to keep versus background - instead of a sloppy blob.
5. Colors are matched to the original boundary, then the crop is blended in with multiband pyramid blending rather than a plain feather.

Where people get burned: the biggest lever is repair_mode. Detail (0.28), Repair (0.42), and Reconstruct (0.58) are fixed denoise presets; Custom (the default, 0.4) uses custom_denoise. Higher denoise reconstructs more aggressively but changes identity and facial geometry - on a character you care about, start at Detail and work up. The other knobs that matter: face_selection (all, largest, largest_n) and max_faces for crowd shots, and detection_quality set to maximum, which runs tiled detection and is what finds the small background faces the full-image pass misses.

Wiring follows the ADetailer-shaped habit: image in, a generation model/vae/positive/negative that matches the source image's style, plus the face_detector connection from the YOLO loader node (or external bboxes, which skip YOLO entirely). face_parser is technically optional, but don't skip it - without it you fall back to SAM/landmark/geometry masks and lose the precise part masks. Outputs: fixed_image goes to your Save Image node; the other four (original_face_crops, processed_face_crops, face_mask, debug_preview) exist so you can see exactly what changed and why.

Install is the pack's shared setup: in ComfyUI Manager search for Ultimate Face Fix, hit Install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Merserk/ComfyUI-Ultimate-Face-Fix.git
cd ComfyUI-Ultimate-Face-Fix
pip install -r requirements.txt
python scripts/prepare_models.py --comfy-root ../..

That last step downloads, verifies, and prepares the three required models (YOLO detector, SegFace parser, MediaPipe landmarks) - it can take a minute on first run. It pins ultralytics==8.4.102, which is the right call given Ultralytics' supply-chain incident in late 2024.

This pack is young - one launch thread in July 2026 - so treat the defaults as a starting point, not doctrine. The early community signal is genuinely good: a tester who ran it for a couple days called it "much better than what I have been using for the last 2 years" and migrated their face repair onto it. The one thing it doesn't do is hold identity across the pass by itself; people pair it with an identity LoRA (works fine - load the LoRA before the model input) and keep denoise under about 0.5.