FusionCow

/

Sulphur-3

Like 609

This model is currently in early access for people who donated to the project. It'll be generaly availibity soon.

Why this long exclusivity period? We don’t even know what it’s worth.

Why this long exclusivity period? We don’t even know what it’s worth.

Long? Its been less than a week, relax. People are working on reducing system requirements, simplifying a few things and just testing things out. If it had been released in its initial state, half of you wouldnt be able to run it since it initially struggled even on16gb vram cards and 64gb memory.

you mentioned "...early access for people who donated to the project..." as in capital? are there capital requirements maybe that might help someone gain access?
my idea for this is not to produce content but use it as a comparison system to vet content submitted images to my social app project where certain types of images are not permitted. since the model has been trained and refined for the specific types content; it might actually work to vet (approve or deny) submitted content but it's all experimental and not sure if it will do what i need it to do but i want to try it. i can build the logic to run a comparison but of course i need access to it.
anyway; no hurry but just wondering.

you mentioned "...early access for people who donated to the project..." as in capital? are there capital requirements maybe that might help someone gain access?
my idea for this is not to produce content but use it as a comparison system to vet content submitted images to my social app project where certain types of images are not permitted. since the model has been trained and refined for the specific types content; it might actually work to vet (approve or deny) submitted content but it's all experimental and not sure if it will do what i need it to do but i want to try it. i can build the logic to run a comparison but this model is the only thing i am aware of trained up for this kind of thing. accuracy has to be worked out but i'm confident this is a worthwhile approach.

anyway; no hurry but just wondering.

Not on the team so can't speak to the donation/early access side, that's a question for the devs. But heads up on the moderation idea: Sulphur is a generative video model, not a classifier. It makes content, it doesn't output an approve/deny score. You can technically use a diffusion model as a zero-shot classifier by comparing denoising loss across prompts, but that's dozens of forward passes per image on a model that barely fits on 16GB cards. For every upload to a social app that's going to be painfully slow and expensive, and there's no reason to think it'd be accurate.

you mentioned "...early access for people who donated to the project..." as in capital? are there capital requirements maybe that might help someone gain access?
my idea for this is not to produce content but use it as a comparison system to vet content submitted images to my social app project where certain types of images are not permitted. since the model has been trained and refined for the specific types content; it might actually work to vet (approve or deny) submitted content but it's all experimental and not sure if it will do what i need it to do but i want to try it. i can build the logic to run a comparison but this model is the only thing i am aware of trained up for this kind of thing. accuracy has to be worked out but i'm confident this is a worthwhile approach.

anyway; no hurry but just wondering.

HF double posted my reply lol

yes; thanks for that... almost immediately after i posted i got to work on something like that which should work out. going at it from the original ask; it would have been more work that wasn't needed and turns out there are a variety of these types of systems that already exist so should workout. if accuracy was a thing i can just build in more so i won't be needing access to this. thanks for the reply by the way!

the wait of a century.....