Workflow Overview

This workflow, titled “SUPIR-8K Wallpaper-Level Upscaling,” aims to upscale low-resolution input images into high-quality, wallpaper-grade images (e.g., 8K resolution). It achieves this through image pre-scaling and SUPIR (Super-Resolution Upscaling) technology, while also allowing comparison between the original and upscaled images. The generation process takes approximately 140 seconds, making it suitable for tasks like art creation, wallpaper design, or image restoration.

Core Models

1. SUPIR-v0F_fp16.safetensors

   • Function：SUPIR is a super-resolution model based on diffusion techniques, designed to enhance low-resolution images into high-quality, detailed outputs.

   • Source：Downloadable from official model repositories or communities (e.g., Hugging Face), to be placed in ComfyUI/models/checkpoints.

2. juggernautXL_v8Rundiffusion_V8

   • Function：A Stable Diffusion variant serving as the base model for SUPIR, providing latent representations to assist in detail generation and restoration during upscaling.

   • Source：Available from Stable Diffusion communities (e.g., Civitai), to be placed in ComfyUI/models/checkpoints.

Component Explanation

1. LoadImage

   • Purpose：Loads a user-provided input image (e.g., PNG or JPG).

   • Function：Outputs an image (IMAGE) and an optional mask (MASK). Here, the image is sent to scaling and upscaling nodes.

   • Installation：Core ComfyUI node, no additional setup required.

2. DF_Image_scale_to_side

   • Purpose：Scales the input image to a specified size (e.g., longest side 1024 pixels).

   • Function：Ensures the image meets size requirements for subsequent processing, using methods like “nearest-exact.” Configured to 1024 pixels, longest side, nearest-exact interpolation.

   • Installation：Requires the ComfyUI-DF-Nodes library, installable via ComfyUI Manager (search “DF Nodes”) or manually from GitHub (https://github.com/DesignerDF/ComfyUI-DF-Nodes) into ComfyUI/custom_nodes.

3. SUPIR_Upscale

   • Purpose：Performs super-resolution upscaling on the scaled image.

   • Function：Offers adjustable parameters (e.g., scale factor, sampling steps, prompts), set here to 2x upscaling, 45 steps, positive prompt “high quality, detailed,” negative prompt “bad quality, blurry, messy.”

   • Installation：Requires SUPIR nodes, installable via ComfyUI Manager (search “SUPIR”) or from GitHub (official SUPIR repository).

   • Dependencies：Requires SUPIR model (SUPIR-v0F_fp16.safetensors) and base model (juggernautXL_v8Rundiffusion_V8).

4. SaveImage

   • Purpose：Saves the upscaled image locally.

   • Function：Receives the upscaled image and saves it with a default prefix “ComfyUI.”

   • Installation：Core ComfyUI node, no additional setup required.

5. Image Comparer (rgthree)

   • Purpose：Compares the original and upscaled images.

   • Function：Displays both images in “Slide” mode for easy visual comparison of the upscaling effect.

   • Installation：Requires the rgthree-comfy library, installable via ComfyUI Manager (search “rgthree”) or from GitHub (https://github.com/rgthree/rgthree-comfy).

Workflow Structure

The workflow is divided into the following logical groups:

1. Image Input and Preprocessing Group

   • Nodes：LoadImage → DF_Image_scale_to_side

   • Role：Loads the input image and scales it to a suitable size (longest side 1024 pixels).

   • Input Parameters：Image file path (e.g., QQ截图20240821173916.png).

   • Output：Scaled image (IMAGE).

2. Image Upscaling Group

   • Node：SUPIR_Upscale

   • Role：Uses the SUPIR model to upscale the scaled image to high quality.

   • Input Parameters：Scaled image, positive prompt “high quality, detailed,” negative prompt “bad quality, blurry, messy,” 2x scale, 45 steps, etc.

   • Output：Upscaled high-quality image (IMAGE).

3. Output and Comparison Group

   • Nodes：SaveImage and Image Comparer (rgthree)

   • Role：Saves the upscaled image and compares it with the original.

   • Input Parameters：Original image (from LoadImage), upscaled image (from SUPIR_Upscale).

   • Output：Saved image file and visual comparison interface.

Inputs and Outputs

• Expected Inputs：

  • Image file: Any resolution PNG/JPG (e.g., QQ截图20240821173916.png).

  • Optional prompts: Positive (“high quality, detailed”) and negative (“bad quality, blurry, messy”).

  • Scaling parameters: Longest side 1024 pixels (adjustable).

  • Upscaling parameters: 2x scale, 45 steps (adjustable).

• Final Outputs：

  • High-quality upscaled image (saved as PNG with “ComfyUI” prefix).

  • Visual comparison of original and upscaled images (via Image Comparer).

Notes and Tips

1. Resource Requirements：SUPIR upscaling demands significant GPU power; at least 8GB VRAM is recommended. For lower VRAM, reduce scale factor or use fp16 models.

2. Model Installation：Ensure SUPIR-v0F_fp16.safetensors and juggernautXL_v8Rundiffusion_V8 are in ComfyUI/models/checkpoints, or errors will occur.

3. Performance Optimization：If generation takes too long, reduce sampling steps (e.g., from 45 to 30) or lower target resolution.

4. Compatibility：Verify all third-party nodes (DF_Image_scale_to_side, Image Comparer) are installed, or the workflow will fail.

5. Troubleshooting：If upscaling results are suboptimal, tweak prompts or use FLUX tile upscaling as suggested by the author (see workflow title note).