FaceCanon — Consistent Faces at Any Resolution (ComfyUI)

Normalize → Detail → Inverse Composite.
FaceCanon scales a detected face to a canonical pixel size, lets you run your favorite face detailer at that sweet spot, then maps the result back into the original image with seamless blending. The payoff is consistent face style no matter the input resolution or framing.

🔍 Before vs After Comparison

Here's a side-by-side comparison showing the difference FaceCanon makes at 768x768 resolution when combined with FaceDetailer:

Notice how much the quality and artstyle of the the bottom-left render has deteriorated

✨ What it does

• Canonical face scale: Locks the face to a target pixel size (e.g. 256 px) before enhancement.
• Works with any detailer: Drop in FaceDetailer, CodeFormer, GPEN, etc.
• Seamless paste-back: Inverse transform + feather or Poisson blending.
• Multi-face aware: Choose largest face or a specific index.
• Resolution-agnostic: Stable results across 1:1, portrait, landscape, SDXL, FLUX, etc.

🎯 When to Use FaceCanon

Most Useful For:

LoRA Model Mixers 🎨 - If you blend multiple artstyle LoRA models, FaceCanon is essential. Without it, face enhancement can create unstable, inconsistent artstyles as different models interpret faces differently at varying resolutions.

Still Helpful For:

Single Model Users 📸 - Even with one consistent model, FaceCanon improves face quality consistency across different input resolutions and framings. The improvement is noticeable but less dramatic than with model mixing.

---

⚙️ Requirements

• ComfyUI
• ComfyUI-Impact-Pack (for Ultralytics/ONNX detectors)

We do not bundle detector weights. See Models below.

---

📦 Install

cd ComfyUI/custom_nodes
git clone https://github.com/SiggEye/FaceCanon.git
pip install -r FaceCanon/requirements.txt
# restart ComfyUI

🧠 Models (Detectors)

FaceCanon requires a face detection model that you must download yourself.

Download Face Detection Model

Recommended model: face_yolov8n.pt

Installation:

1. Download the model file from the link above
2. Place it in your ComfyUI models directory:

   ComfyUI/models/ultralytics/bbox/face_yolov8n.pt
   

Setup in ComfyUI

1. In your workflow, create an UltralyticsDetectorProvider node
2. Select face_yolov8n.pt as the model
3. Connect the BBOX_DETECTOR output (green) to FaceCanon's bbox_detector input

Note: You only need the BBOX_DETECTOR output (green). The SEGM_DETECTOR output (red) can be left unused.

🧩 Nodes

FaceScaleNormalize

Detects a face and warps the source image into a canonical canvas.

Inputs:

• image - Input image tensor
• bbox_detector - Face detector from Impact-Pack (required)
• base_size - Canvas size (default: 1024, divisible by 16)
• target_face_px - Target face size in pixels (default: 256)
• face_selection - "largest" or "index" (default: "largest")
• face_index - Face index when using "index" selection (default: 0)
• center_face - Center face on canvas (default: true)
• oversize_strategy - "downscale_to_fit" or "center_crop"
• pad_mode - "edge" | "reflect" | "constant" (+ pad_color if constant)
• pad_color - RGB color for constant padding (default: "128,128,128")

Outputs:

• normalized_image - Image normalized to canonical dimensions
• transform_meta - JSON metadata for inverse transform

FaceScaleDenormalizeComposite

Maps the edited face back into the original image with seamless blending.

Inputs:

• original_image - The original input image
• edited_normalized - Enhanced face from your detailer/enhancer
• transform_meta - Transform metadata from FaceScaleNormalize
• mask_expand_px - Pixels to expand face mask (default: 32)
• feather_px - Gaussian blur radius for blending (default: 32)
• blend_mode - "feather" or "poisson" (default: "feather")
• use_ellipse_mask - Use elliptical mask for natural blending (default: true)

Output:

• composited_image - Final result with enhanced face blended back

🔄 Basic Workflow

[Image] → UltralyticsDetectorProvider → FaceScaleNormalize
                                   ├─→ (normalized_image) → FaceDetailer → FaceScaleDenormalizeComposite → [Final]
                                   └─→ (transform_meta) ─────────────────────────────────────────────────┘

Quick Setup:

1. Drop UltralyticsDetectorProvider, select your face model
2. Set base_size (e.g., 1024) and target_face_px (e.g., 256) on FaceScaleNormalize
3. Pipe normalized_image into your FaceDetailer (or any enhancer)
4. Feed edited_normalized + transform_meta + original_image to FaceScaleDenormalizeComposite
5. Tweak feather_px or switch to poisson if you see seams

📸 Example Workflow

Here's a visual comparison of a ComfyUI FaceDetailer setup without (top) and with (bottom) FaceCanon:

✅ Recommended Defaults

| Parameter | Value | Notes | |-----------|-------|-------| | base_size | 1024 | Divisible by 16, good for SD models | | target_face_px | 256 | Optimal face size for most detailers | | pad_mode | "edge" | Good for uniform backgrounds | | blend_mode | "feather" | Fast, usually looks great | | feather_px | 32 | Good balance of sharpness/blending | | mask_expand_px | 32 | Prevents visible edges |

💡 Pro Tip: If your best results happen when faces are ~200px on a 1024×1024 canvas, try: base_size: 1024, target_face_px: 200

👥 Multi-Face Support

Quick single face: Set face_selection: "largest"

Specific faces: Set face_selection: "index" and run separate branches with:

• face_index: 0 (first detected face)
• face_index: 1 (second detected face)
• etc.

🧐 FAQ

Q: Isn't this just setting guide_size in FaceDetailer?
A: No! Guide/max sizes don't guarantee consistent face pixel sizes across arbitrary inputs.

🧪 Roadmap

• [ ] Tiny demo workflows for SDXL / FLUX
• [x] Before→after comparison in README
• [ ] Before→after GIFs in README

🏷️ License

MIT — See LICENSE file for details.