1. Workflow Overview

• Function: Generates high-quality images of ancient-style beauties holding umbrellas, with face/hand refinement, face-swapping, and SUPIR upscaling.

• Goal: Combine LoRA models and SUPIR to enhance stylized portrait generation.

• Key Models:

  • Stable Diffusion 1.5 (Base: majicMIX realistic_v7).

  • SUPIR-v0F (Fast upscaling model).

  • ControlNet (YOLOv8 for face/hand detection).

  • LoRA Stack (e.g., Ancient Beauty LoRA, Hand Fix LoRA).

2. Node Breakdown

Critical Nodes:

1. CheckpointLoaderSimple: Loads base model (e.g., majicMIX realistic).

2. CR LoRA Stack: Dynamically merges multiple LoRAs (costume, hand fix, detail enhancer).

3. KSampler: Controls generation (sampler: dpmpp_2m, steps 25-50).

4. FaceDetailer/HandDetailer: Repairs faces/hands using YOLOv8 detectors.

5. SUPIR_sample: Upscales images with tile-based processing.

6. ReActorFaceSwap: Swaps faces (requires source portrait).

Dependencies:

• SUPIR Models: Download SUPIR-v0F_fp16.safetensors to models/supir.

• UltralyticsDetector: Install comfyui-controlnet-aux for bbox/face_yolov8m.

3. Workflow Structure

Group Logic:

1. Model Loading:

   • Input: Base model, LoRAs, VAE.

   • Output: Loaded model + CLIP encoder.

2. Prompts:

   • Input: Positive/Negative prompts (e.g., "hanfu, umbrella").

   • Output: Conditioned text tensors.

3. Image Generation:

   • Input: Empty latent (512x768), text conditions, seed.

   • Output: Raw generated image.

4. Refinement & Upscale:

   • Input: Raw image, face/hand detectors.

   • Output: 2K upscaled image via SUPIR.

4. Inputs & Outputs

Inputs:

• Required: Positive/Negative prompts, seed (randomizable).

• Optional: Face-swap source image, SUPIR scale (default 4x), CFG (default 4).

Outputs:

• Final image: 2K PNG with metadata.

5. Notes

1. Hardware: ≥8GB VRAM; SUPIR needs extra memory.

2. Common Errors:

   • Missing SUPIR model: Check models/supir.

   • Face repair fails: Adjust FaceDetailer threshold (default 0.55).

3. Optimization:

   • Reduce encoder_tile_size (e.g., 512) for lower VRAM usage.

   • Enable fp8_unet if GPU supports FP8.