Workflow Overview

This workflow is a sophisticated image generation and enhancement pipeline based on Stable Diffusion, designed to create high-quality Chinese "Guochao" style illustrations. Its main purposes are:

1. Generating initial images using Stable Diffusion.

2. Enhancing details via image blending (ImageBlend) and depth mapping (DepthAnything V2).

3. Refining images with ControlNet (depth-based) and multiple sampling passes (KSampler).

4. Upscaling the final image to a higher resolution using RealESRGAN and UltimateSDUpscale.

Core Models

• Stable Diffusion: Core generation model, loaded with “锦绣芳华——国潮插画风_v1.0.safetensors” for Guochao-style images.

• LoRA (国潮-插画艺术_v1.0.safetensors): Enhances the Guochao illustration style.

• CLIP: Loaded from “CheckpointLoaderSimple” for text prompt processing.

• VAE (vae-ft-mse-840000-ema-pruned.ckpt): Encodes/decodes images for generation optimization.

• ControlNet (control_v11f1p_sd15_depth.pth): Controls generation using depth maps.

• DepthAnything V2 (depth_anything_v2_vitl_fp32.safetensors): Generates depth maps for enhanced 3D effects.

• RealESRGAN (RealESRGAN_x2.pth and RealESRGAN_x4plus_anime_6B.pth): Super-resolution models for upscaling.

Component Explanation

1. CheckpointLoaderSimple: Loads Stable Diffusion, CLIP, and VAE.

   • Installation: Default ComfyUI node.

2. KSamplerAdvanced: Advanced sampler for generating/refining latent images.

   • Installation: Default ComfyUI node.

3. EmptyLatentImage: Creates an empty latent image.

   • Installation: Default ComfyUI node.

4. CLIP Positive-Negative (WLSH): Processes positive/negative prompts.

   • Installation: Install via ComfyUI Manager (WLSH custom nodes).

5. VAEDecode: Decodes latent images into visible images.

   • Installation: Default ComfyUI node.

6. ImageBlend: Blends two images.

   • Installation: Default ComfyUI node.

7. DepthAnything_V2: Generates depth maps.

   • Installation: Install via ComfyUI Manager; model from Hugging Face.

8. ControlNetLoader and ControlNetApply: Load and apply ControlNet.

   • Installation: Default nodes; model from Civitai or Hugging Face.

9. UpscaleModelLoader and ImageUpscaleWithModel: Load and apply upscaling models.

   • Installation: Default nodes; models from GitHub or RealESRGAN official sources.

10. HD UltimateSDUpscale: Advanced super-resolution upscaling.

    • Installation: Install via ComfyUI Manager (UltimateSDUpscale plugin).

11. AV_VAELoader: Loads an external VAE model.

    • Installation: Install via ComfyUI Manager (AnimateVision plugin).

Workflow Structure

1. Initial Image Generation Group

   • Nodes: EmptyLatentImage → KSamplerAdvanced → VAEDecode

   • Role: Generates two initial images (512x288 and 512x1024).

   • Inputs: Prompts (e.g., “Guochao(style), distant mountains”), seed, sampling steps.

   • Outputs: Two initial images.

2. Image Blending and Depth Enhancement Group

   • Nodes: ImageBlend → DepthAnything_V2 → VAEEncode

   • Role: Blends two initial images and generates a depth map, then encodes back to latent space.

   • Inputs: Two initial images, blend factor (0.38).

   • Outputs: Blended latent image and depth map.

3. ControlNet Refinement Group

   • Nodes: ControlNetLoader → ControlNetApply → KSamplerAdvanced

   • Role: Refines the image using a depth-based ControlNet.

   • Inputs: Depth map, prompts, ControlNet strength (0.7).

   • Outputs: Refined latent image.

4. Super-Resolution Upscaling Group

   • Nodes: UpscaleModelLoader → ImageUpscaleWithModel → HD UltimateSDUpscale

   • Role: Progressively upscales the image to a higher resolution.

   • Inputs: Refined image, upscaling models.

   • Outputs: Final high-resolution image.

Inputs and Outputs

• Inputs:

  • Positive prompts: e.g., “Main building, clouds, sky, mountains, guochaochahua”.

  • Negative prompts: e.g., “mankind, monochrome”.

  • Resolution: Initial 512x288 and 512x1024, upscaled to higher resolution.

  • Seed: Random or specified.

• Outputs: A high-quality Guochao-style illustration image (PNG format after upscaling).

Notes and Considerations

• Errors: Ensure correct model paths to avoid “model not found” issues.

• Performance: Use FP16 precision to reduce memory usage.

• Compatibility: WLSH and UltimateSDUpscale require the latest ComfyUI version.

• Resources: Recommend at least 12GB GPU memory; CPU execution may be slow.