[Batched Generators] This PR adds generators that are useful to make batched generation fully reproducible #1718
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
|
Tests are flaky I think. |
anton-l
left a comment
There was a problem hiding this comment.
LGTM! Not sure about updating the types and docs for incompatible pipelines like DDPM and K-Diffusion though, but depends on how quickly we add the feature there afterwards :)
pcuenca
left a comment
There was a problem hiding this comment.
This should be very useful, thanks a lot!
| >>> images | ||
| ``` | ||
|
|
||
|  |
There was a problem hiding this comment.
Lol, that doesn't look like the style of Vermeer! Was he removed from the training set of 1.5? (Nothing to change here, just a comment; I was curious about the double eyes).
| # TODO(Pedro) - not sure why, but not at all reproducible at the moment it seems | ||
| # make sure that batched and non-batched is identical |
There was a problem hiding this comment.
This is possibly because of pytorch/pytorch#84039 (comment). We can't really support more than 1 image on mps yet.
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
…batched generation fully reproducible (huggingface#1718) * [Batched Generators] all batched generators * up * up * up * up * up * up * up * up * up * up * up * up * up * up * up * up * hey * up again * fix tests * Apply suggestions from code review Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * correct tests Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
|
For a batched generation, how do I use the generator? I passed in a list of generators (same length as the number of images per prompt) with the same manual seed number but the first image differs from the other images (other images 2, 3, 4 are the same) in a batch? If I put in just one generator, then all the images in the batch are different. But using the same seed results in the same batch of images. How do I do per image seed? Say if I use seed 42, 42, 42, 42 I expect all the images returned are the same image. More findings, it seems that txt2img pipeline is fine but not img2img. This is Stable Diffusion btw. |
|
Hey @xiankgx, Could you please open a new issue with a reproducible code snippet? |
|
@xiankgx @patrickvonplaten i monkeypatched it with this solution https://gist.github.com/kylemcdonald/208987e0b6c562886ae2cf42fb5fb743 use it like this: from fixed_noise import fix_noise
...
fix_noise(pipe)this will use the same random noise for all images in the batch. |
…batched generation fully reproducible (huggingface#1718) * [Batched Generators] all batched generators * up * up * up * up * up * up * up * up * up * up * up * up * up * up * up * up * hey * up again * fix tests * Apply suggestions from code review Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * correct tests Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
This PR adds two nice features: