1. Workflow Overview

This workflow, "Game Design Focus: Draft Control + Style Transfer + WebUI Restoration", is optimized for game character design, featuring:

• Sketch Control: Uses ControlNet (xinsir_controlnet_tile_sdxl and diffusers_xl_depth_mid) for precise structure.

• Style Transfer: Applies IPAdapter to transfer reference styles.

• SDXL Enhancement: Leverages Jianzhen SDXL model with LoRAs (e.g., Wheelchair Fashion).

• Upscaling: Combines UltimateSDUpscale and SwinIR_4x for HD output.

Core Models:

• Jianzhen SDXL Wheelchair: Base model (realistic style).

• IPAdapter-MoE: Multi-style fusion adapter.

• ControlNet: xinsir_controlnet_tile_sdxl (sketch), diffusers_xl_depth_mid (depth).

2. Key Nodes

1. IPAdapterUnifiedLoader (Node 564)

   • Function: Loads IPAdapter for multi-style fusion (e.g., PLUS-high strength).

   • Install: Via ComfyUI Manager (ComfyUI_IPAdapter_plus).

2. ControlNetApplyAdvanced (Nodes 427/430)

   • Function: Applies sketch/depth control (weights: sketch 0.25, depth 0.8).

   • Dependency: comfyui_controlnet_aux package.

3. UltimateSDUpscale (Node 633)

   • Function: Tile-based upscaling with SwinIR_4x (tile size 512x512).

4. BNK_CLIPTextEncodeAdvanced (Nodes 645/646)

   • Function: Advanced text encoder with negative prompts (e.g., "bad anatomy").

5. Jianzhen SDXL LoRA (Node 657)

   • Function: Enhances details (e.g., armor textures) via Wheelchair Fashion LoRA.

Special Dependencies:

• Model Files:

  • Jianzhen SDXL Wheelchair.safetensors: Download to models/checkpoints.

  • xinsir_controlnet_tile_sdxl: Get from CivitAI, place in models/controlnet.

3. Workflow Structure (Groups)

1. Group 1: Model Loading (Nodes 542/546/657)

   • Loads base model, VAE, and LoRAs.

2. Group 2: Draft Input & Preprocess (Nodes 655/560)

   • Input: Sketch or design draft (e.g., 02.png).

   • Resize: ImageResizeKJ adjusts resolution to 2048x2048.

3. Group 3: ControlNet (Nodes 427/430/434/431)

   • Sketch: LineartStandardPreprocessor extracts clean lines.

   • Depth: MiDaS-DepthMapPreprocessor adds 3D perception.

4. Group 4: IPAdapter Style Transfer (Nodes 564/621/565)

   • Input reference image (e.g., 433383010-xxx.png) for style fusion.

5. Group 5: Generate & Upscale (Nodes 618/633)

   • Generation: KSampler //Inspire with dpmpp_2m (38 steps).

   • Upscale: UltimateSDUpscale repairs details tile-by-tile.

6. Group 6: Output & Compare (Nodes 656/318)

   • Saves final image and provides slider comparison (original vs. generated).

4. Inputs & Outputs

Inputs:

• Draft Image: Recommended 768x1280 (text-to-image) or 2048x2048 (image-to-image).

• Prompts: Detailed descriptions (e.g., "futuristic soldier, grey armor").

• Reference Image: For style transfer (clean background preferred).

Outputs:

• HD character design (saved as PNG).

• Slider comparison tool.

5. Notes

1. Optimization:

   • Reduce UltimateSDUpscale tile size (e.g., 384x384) for low VRAM.

   • Enable --xformers for faster inference.

2. Troubleshooting:

   • Sketch control fails → Check ControlNet model paths.

   • Style deviation → Adjust IPAdapter weights (default 0.1-0.8).

3. Compatibility:

   • Requires ComfyUI ≥ 0.3.30 and PyTorch 2.1+.

   • BNK_CLIPTextEncodeAdvanced needs ComfyUI_ADV_CLIP_emb.