ComfyUI-Flux-FaceIR

ComfyUI Flux FaceIR

ComfyUI extension for FLUX FaceIR face restoration. It supports both:

• aligned face restoration from pre-cropped face patches
• full-image restoration from original images with face detection, alignment, parameter export, and paste-back

Main training/inference repository: cosmicrealm/flux-restoration

Workflows

This extension ships with two workflows:

• workflows/aligned_face_restore.json
• workflows/full_image_restore.json

aligned_face_restore.json is for already aligned face crops.

full_image_restore.json is the whole-image workflow:

1. load one degraded full image
2. load reference full images
3. detect and align faces with RetinaFace
4. restore the aligned degraded face with FaceIR
5. export alignment parameters
6. paste the restored face back to the original degraded image

The detect-and-align node returns both:

• align_params: structured affine/paste-back parameters for downstream nodes
• align_params_json: readable JSON text for debugging or export

Screenshots

Aligned workflow:

Whole-image workflow:

Installation

ComfyUI Manager / Registry

Search for Flux FaceIR in ComfyUI Manager, or install it with the Comfy CLI:

comfy node install flux-faceir

The two bundled workflows will then be available from ComfyUI's workflow template browser. Model weights are installed separately; see Required Models.

Manual installation

Clone this repository into ComfyUI/custom_nodes, then install dependencies:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/cosmicrealm/ComfyUI-Flux-FaceIR.git ComfyUI-Flux-FaceIR
cd ComfyUI-Flux-FaceIR
python install.py

If you downloaded a zip or received this extension bundled with another package, extract or copy it so the final directory is:

ComfyUI/custom_nodes/ComfyUI-Flux-FaceIR

Do not leave the directory named ComfyUI-Flux-FaceIR-main. Fully restart ComfyUI after installation.

Required Models

Recommended model layout:

ComfyUI/models/
  diffusion_models/
    FLUX.2-klein-base-4B/
      flux-2-klein-base-4b.safetensors
  text_encoders/
    qwen_3_4b.safetensors
  vae/
    flux2-vae.safetensors
  loras/
    lora_weights.safetensors
  face_detectors/
    retinaface_r34.pth

Download links:

• FLUX base model: flux-2-klein-base-4b.safetensors
• Qwen text encoder: qwen_3_4b.safetensors
• FLUX VAE: flux2-vae.safetensors
• FaceIR LoRA: lora_weights.safetensors
• RetinaFace detector: retinaface_r34.pth

Example download commands:

mkdir -p ComfyUI/models/face_detectors
wget -O ComfyUI/models/face_detectors/retinaface_r34.pth \
  https://github.com/yakhyo/retinaface-pytorch/releases/download/v0.0.1/retinaface_r34.pth

mkdir -p ComfyUI/models/loras
wget -O ComfyUI/models/loras/lora_weights.safetensors \
  https://huggingface.co/zhangjinyang/flux-restoration/resolve/main/pretrained_models/lora_weights.safetensors

Node Overview

Core nodes:

• Flux FaceIR Apply LoRA
• Flux FaceIR Load RetinaFace
• Flux FaceIR Detect And Align Face
• Flux FaceIR Restore Face
• Flux FaceIR Paste Restored Face

Recommended graph:

1. Load base FLUX components with UNETLoader, CLIPLoader, and VAELoader.
2. Apply lora_weights.safetensors with Flux FaceIR Apply LoRA.
3. If your inputs are already aligned: use Flux FaceIR Restore Face directly.
4. If your inputs are full images: use Flux FaceIR Load RetinaFace -> Flux FaceIR Detect And Align Face -> Flux FaceIR Restore Face -> Flux FaceIR Paste Restored Face.

The bundled whole-image workflow is configured for:

• detector weights: retinaface_r34.pth

Notes

• This extension does not download weights automatically.
• Base model loading follows standard ComfyUI loaders.
• The bundled workflow targets the non-quantized flux-2-klein-base-4b.safetensors, not the FP8 variant.
• Flux FaceIR Restore Face supports blind restoration and reference-guided restoration with up to three references.
• The full-image workflow only pastes the restored face back into the degraded source image. Reference images are used only for guidance.
• crop_scale, crop_shift_y, detector thresholds, and resize settings are exposed on the detect-and-align node, so you can widen the crop to include more forehead, chin, or shoulder context when needed.