Supir image restoration and upscaling guide for Comfy

Supir helps you keep control and consistency in images while upscaling or restoring. It’s a simple Comfy workflow you can download (see bottom of page) and drag & drop into your local Comfy or cloud solution. 

Go into the manager and select “Install Missing Custom nodes”. Select any nodes that appear in the list and press Install. Then press restart and refresh page.

Select an SDXL model. I used a lightning model to increase speed. Remember to adapt your CFG to your model (see further down). Try various lightning models to improve your results.

https://civitai.com/models/139562?modelVersionId=361593

Select your SUPIR model. Recommended is the Q version.

You can download from the manager by pressing “Install Models”

Search for Supir and press Install on the two safetensor ones Q and F.

OR

Download from here: https://huggingface.co/Kijai/SUPIR_pruned/tree/main

And place in /ComfyUI/models/checkpoints/

SUPIR-v0Q

Default training settings with paper. High generalization and high image quality in most cases.

SUPIR-v0F

Training with light degradation settings. Stage1 encoder of SUPIR-v0F remains more details when facing light degradations.

Fp8_unet: Set to true if you have low end gpu. False for high end gpu.

High_vram: Set to true if you have high end gpu. False for low end gpu.

Select your upscale factor. An image in 1024x1024 in factor 2 will be 2 times larger, ie. 2048x2048.

Select your CFG based on the model you use (see Civitai page). Most lightning models use low CFG around 1. Regular SDXL models work well around 3-7 CFG.

Select your Supir Control scale based on how much you want to retain from your input image. Control scale value between 0-1. Anything over 1 will look overcooked. Higher value = more consistent to the original image.

Write your prompt. The better you describe your image, the better Supir will work for you.

The first box is the positive prompt and the bottom box is the negative prompt. 

Seed randomized or fixed depending on your needs. A fixed seed is set to keep generating the same image, ie. when testing to finetuning your settings. Leave randomized default.

Steps is dependant on your model. Lightning models work well 8 or below. Regular SDXL models around 20-40

Sampler, try Restore and TiledRestore. Tiledrestore will work in tiles, generating small square pieces at a time. If control scale is low, tiledrestore will produce your prompt in multiple squares. If lowering control scale, try to change to non-tile sampler.

Load your image in the load image node. Any size works. 

The image comparer slides between before and after images.