Feature extraction from Unet of Stable Diffusion

[danushkv]

Thanks to the HF community for this wonderful library. I'm currently working on a project where we are interested in looking into the intermediate features(feature maps from different layers) of the Unet in Stable diffusion model.

The transformers library has the option of output_hidden_states to extract the intermediate features from a transformer model. Similarly, is there an option to do so from the StableDiffusionPipeline ? I couldn't find an option for this, but please correct me if I'm wrong here.

If there's isn't an option, we would like to have one. I would be interested in working on this as a project also.

[sayakpaul]

Since this feature is not a widely requested one, we are currently not prioritizing it.

@patrickvonplaten @williamberman what are your thoughts?

[danushkv]

@sayakpaul thanks for the reply. There are few projects that focus on extracting features from SD models (especially from the Unet). I would like to work on this and submit it as a PR. If this still satisfies as a need (maybe in the future) you can definitely merge it.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[williamberman]

I understand the use case but I think we might need to see a few more requests for it before adding? I think the needed code is small enough if someone wants to extract intermediate results for a one off project. Perfectly reasonable feature though if more people request it

[aisagarw]

Do we have this functionality anywhere now? I too was looking at extracting features from the Unet of Stable Diffusion..

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.