https://github.com/haoming02/sd-webui-resharpen

An Extension for Automatic1111 Webui that increases/decreases the details of images
https://github.com/haoming02/sd-webui-resharpen

stable-diffusion-webui stable-diffusion-webui-plugin

Last synced: 7 months ago
JSON representation

An Extension for Automatic1111 Webui that increases/decreases the details of images

• Host: GitHub
• URL: https://github.com/haoming02/sd-webui-resharpen
• Owner: Haoming02
• License: mit
• Created: 2023-11-14T02:39:02.000Z (almost 2 years ago)
• Default Branch: main
• Last Pushed: 2024-09-02T04:05:44.000Z (about 1 year ago)
• Last Synced: 2025-02-28T14:25:17.198Z (8 months ago)
• Topics: stable-diffusion-webui, stable-diffusion-webui-plugin
• Language: Python
• Homepage:
• Size: 747 KB
• Stars: 82
• Watchers: 1
• Forks: 5
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# SD Webui ReSharpen

This is an Extension for the [Automatic1111 Webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui), which allows you to increase or decrease the amount of *details* of the generation during the Stable Diffusion pipeline.

> **ie.** This is not just a postprocessing filter

## How to Use

There is simply a slider that controls the "sharpness" of the images, defaulting to 0.

- **Positive** values cause the images to be **noisy**

- **Negative** values cause the images to be **blurry**

> Though, don't actually use values too close 1 or -1 as it will become distorted

For `txt2img`, there is also a second slider for the `Hires. Fix` pass.

You can set positive for the normal pass and negative for the hr pass, or vice versa, for some interesting effects.

There is also now a scaling setting, allowing you to adjust the strength throughout the process *(**eg.** lower the sharpness towards the end)*.

> This Extension is disabled during the [ADetailer](https://github.com/Bing-su/adetailer) phase

**Note:** `Ancestral` samplers *(**eg.** `Euler a`)* do **not** work. Tested and confirmed to work for both `Euler` and `DPM++ 2M Karras`.

## Samples

    

        

            Sharpness

            -0.5

            -0.25

            0.0

            0.25

            0.5

        

    

    

        

            Result

            

            

            

            

            

        

    







## How does it work?

I initially wrote this Extension to try speeding up the Diffusion process.

Since the latent noises gradually converges to around 0,

I tried to see if I can just add the delta between the current step and the previous step, perhaps multipled by a scaler,

to the current Tensor, in order to help it reaches 0 faster.

However, after fiddling around for a week, I still couldn't get it to consistently produce an actually better result.

So I gave up on the original idea.

Though in the process, I also found out this "sharpening" interaction. If I had to guess,

since the Diffusion is a denoising process, if I add the delta again, then it will overly denoise to create a blurry image;

if I subtract the delta, then it will basically add the noise back to create a noisy image.

## To Do

- [X] Generation InfoText

- [X] Pasting InfoText Support

- [X] X/Y/Z Plot Support

- [X] Scaling Support