comfyanonymous / ComfyUI

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.
https://www.comfy.org/
GNU General Public License v3.0
57.26k stars 6.06k forks source link

Img2Img Denoising is broken, setting it to 1 (full) still impacts the sampling #1077

Open n00mkrad opened 1 year ago

n00mkrad commented 1 year ago

Setting denoise to 1 will still show influence of the input image:

image

Expected behavior: Denoising 1.0 should fully denoise the input and thus the image should have zero impact on the sampling. However, it apparently does not get fully denoised. The output is still influenced by the image instead of giving the same result as passing an empty latent image.

comfyanonymous commented 1 year ago

That's normal. In ComfyUI denoise 1.0 is only equivalent to txt2img when the input latent image is zeros (EmptyLatentImage). When the input image is not zeros it will slightly shift the noise in that direction.

n00mkrad commented 1 year ago

That's normal. In ComfyUI denoise 1.0 is only equivalent to txt2img when the input latent image is zeros (EmptyLatentImage). When the input image is not zeros it will slightly shift the noise in that direction.

But then how can I do img2img that's closer to the input image?

Is there any reason the scale doesn't go from no denoise to full denoise, like in InvokeAI or A1111?