data: publish generated dataset
Publish with a permissive license a dataset of generated images :
- confirm models : various stable diffusion only ? (would it be possible to generate / but not publish blackforestlabs and deepfloyd models ?)
- confirm prompts and saved metadata
- confirm volume : ~ 500 / model / data-domain (for now :
500 * 4 models * (face + non-face) = 4_000 images
) - choose appropriate host (in relation with communication's team) :
-
Zenodo :
➕ = research oriented (do not prevent us from realising on HuggingFace) -
HuggingFace ? :
➖ sovereignty➖ lost in the shuffle as there are many many hosted datasets (we should probably discuss firstly our global hf comm strategy as we dont have any account for now)
-
Zenodo :
- confirm that there are no legal issues (faces, prompts derivated from sfhq + ccaptions, + licenses of models)
Blocked by current refacto of pipelines
Edited by Gaspard Defréville