SUPIR Upscaler¶

This workflow is added mostly for research purposes, it is still in development.

Memory requirements(both VRAM and RAM) are directly related to the input image resolution.

Currently, for macOS runners Diffusion type must be set to fp32.

Low memory mode: reduces the size of processed tiles to 256.

Note

If you have a very small input image and the result is less than 1024 (512 for low memory mode) pixels in width or height, tiles should be disabled.

From ComfyUI-SUPIR repo:

Memory requirements are directly related to the input image resolution.
In my testing I was able to run 512x512 to 1024x1024 with a 10GB 3080 GPU,
and other tests on 24GB GPU to up 3072x3072.

System RAM requirements are also hefty, don't know numbers
but I would guess under 32GB is going to have issues, tested with 64GB.

Examples¶

This Upscaler is still in development stage, results may be get better.

We specifically place one portrait example where results is not perfect.

But for many tests we performed - portrait scaling is shiny compared to older scaling methods.

Image of a classic car:

Jackie Chan portrait:

Shakira: