Make "steps" in x/y plot more efficient

[Vega-KH]

Ancestral samplers produce vastly different images at different step counts, so I commonly use Euler A and X/Y plot to see my favorite seeds at 30, 50, 70, and 90 steps. But for each image, IT STARTS FROM SCRATCH!

Wouldn't it be great if, after 30 steps, the latent space image was saved, and inference was resumed from there? For the above example, it would take 90 steps total, and give me 4 unique images. As it is now, it takes 30+50+70+90=240 steps.

Alternatively, I've considered making an add-on specifically designed for ancestral samplers, but I'd rather see it implemented directly in the X/Y plot, and I don't have the expertise at the moment to code it.

EDIT: This will not work because, as @przemoc explained, Xth step in N step generation != Xth step in X step generation.

[timntorres]

Hmm, can you verify that, all else held constant, the 30th step of a 90-step generation equals the final step of a 30-step generation? I remember people discovering this equality did not exist back when SD was new. But could be different for these samplers you mention?

[Vega-KH]

Hrmmm... I actually have no idea how to verify that. The setting "show progress every N steps" seems to indicate that it's already possible, but it doesn't save the previews. I think I'd have to code a custom solution just to test that question.

[przemoc]

Generally Xth step in M step generation != Xth step in N step generation.
It's more complex than that.

You can even see it yourself using:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts#saving-steps-of-the-sampling-process

For some very high-level understanding you may check following resources:

• How AI Image Generators Work (Stable Diffusion / Dall-E) - Computerphile
• Stable Diffusion in Code (AI Image Generation) - Computerphile
• The Illustrated Stable Diffusion - Jay Alammar

[Vega-KH]

Wow, you are correct, and my idea won't work at all. Thanks for the information.

[Clayton-Brandt]

Where would I put this code? It's supposed to save every step of an image that gets generated, but I don't know where to put it. I found the image in this website: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts#saving-steps-of-the-sampling-process

[Gerschel]

Welcome to the scene.
scripts folder.
./stable-diffusion-webui/scripts/<name it whatever you like>.py