Playground v2.5 Technical Report, SoTA Open-Source Image Gen Model [pdf]

marketing-cdn.playground.com

17 points·aleks_jpg··6 comments

Not open source, based on their license [0], same as Stable Diffusion then. Are there any actually open source SoTA AI that have a normal open source license rather than having use restrictions?

[0] https://huggingface.co/playgroundai/playground-v2.5-1024px-a...

Do you mean not actually open source because the OpenRail license (and this license) requires responsible use restrictions unlike, say, the MIT license, or because this license requires a license for commercial use above certain revenue limits?

If the latter, SDXL and SDXL Lightning is the best you get, and it's likely not far behind Playground v2.5 if you apply a finetune to output the Midjourney / Playground style.

Both; it's not actually an OSI-defined open source license. Licenses like MIT are, since they have no restrictions on usage, commercial or otherwise.

Seems like a big deal that this is open weights. Am I missing something or is this a very serious contender to Stable Diffusion?

It's a great alternative base model to SDXL for the community, yes. But we'll have to see whether the community using open weights sticks to the SDXL ecosystem after SD3 gets released.

They mostly compare it to the base SDXL. There is only one comparison with a community finetune on page 12 (RealStock v2, I suppose the one from Fooocus) and they don't even link to the model. I would like to see more comparisons with community models based on SDXL since they usually generate better images

Reply on news.ycombinator.com