Use image as seed?

[lachlanphillips]

Not to be confused with img2img, I specifically want to use an image as the seed noise rather than the automatically generated noise.

Does anyone know if this is technically feasible? Thanks

Edit: I thought it's worth clarifying. I still want to prompt with an image, but I also want to prompt with a specific noise image as well. I can't bake the noise into the image because I still only have the option of "adjusting the CFG scale" in order to make anything other than the exact image show up, which then introduces the seed noise back in (which I have no control over)

The benefit to this approach would be being able to animate the noise matte ahead of time and get reliable results from the same noise laid out differently.

[5KilosOfCheese]

Normal img2img works fine for this. You just have make a gaussian noise image. I have tested this myself out of curiosity. However you wil need A LOT more processing. You will however end up with a lot of artifacts and noise - from my limited experience. I guess if you clean the noise so that all excess gets removed it'll work?

It'll need it's own workflow, but if you want to see if you can get it to work, just drop a image filtered with gaussian noise to img2img. However you will need to do lots of iterations with something like loopback and low denoise so that the AI doesn't try guess something in to it.

The origin image was 100% noise done quickly on photoshop.

I have actually used sometimes to encourage the AI to fill something with totally different thing. But then inpainting and inpainting models started to work properly along with masking, I stopped bothering to do this.