Question about PartiPrompts about gill HOT 4 CLOSED

Awesome, thanks so much!

from gill.

Thanks, we've added more details here in the README about the structure. It is the same structure as the original PartiPrompts dataset, but we add a column with human annotations indicating whether they preferred the retrieved or generated images.

is the raw pairwise preference data public (with both annotations)

We've uploaded the annotations for the realism/fidelity setting here.

Can we access the exact images that the human annotators saw

Unfortunately we will most likely be unable to release the exact images used because there are some prompts involving humans, and we are also unable to distribute CC3M images. The generated images can be reproduced by running the PartiPrompts through Stable Diffusion v1.5 and the retrieved images can be acquire by using CLIP ViT-L to retrieve the closest image in CC3M (details in Sec 3.3 of the paper).

Hope that helps!

from gill.

I see, thanks @kohjingyu that helps. For the generated images being as close as possible to the ones you'd have likely shown to the users, could you please provide the exact StableDiffusion configuration you used i.e. text guidance, random seed, number of diffusion steps etc.
Thanks again! :)

from gill.

I've uploaded the Stable Diffusion script here. Hope that helps!

from gill.