Active member
- Joined
- Jun 5, 2018
- Messages
- 345
SDXL 0.9 has "leaked". Torrent: https://drive.google.com/file/d/1J-2KhUG7ZvcN6H_-BIoQgdhkShWRU2YZ/view?usp=sharing. HF diffusers already has the code for it: https://github.com/huggingface/diffusers/pull/3859. Official 1.0 release should follow this month.
For those who don't know, it's a text-to-image generator you can run on your own local computer. The most basic mode of operation is you giving it an image caption text, it generating matching images. Their blog post: https://stability-ai.squarespace.com/blog/sdxl-09-stable-diffusion
It trounces the old Stable Diffusion stuff in ability to generate correct-looking things, and unlike DeepFloydIF the release isn't botched (no pessimistic "minimum" VRAM requirements, emphasizing spooky shrinkwrap license with language about enabling others to bypass restrictions, or substituting a really inferior final upscaling stage instead of finishing training). It could unseat SD 1.5 as the default base model.
OTOH the samples I've seen so far are biased towards lame cinematic pablum. But I haven't tried it myself yet, so dunno the actual range of its aesthetics, and lots of people love that kind of stuff anyways.
SDXL is also significant in that it might be the last substantial/expensive-to-train locally-runnable image model to be released for a while - AFAIK there's nothing announced to be upcoming after this. Stability AI must be starting to run low on funds, and Runway and Sberbank were one-offs.
For those who don't know, it's a text-to-image generator you can run on your own local computer. The most basic mode of operation is you giving it an image caption text, it generating matching images. Their blog post: https://stability-ai.squarespace.com/blog/sdxl-09-stable-diffusion
It trounces the old Stable Diffusion stuff in ability to generate correct-looking things, and unlike DeepFloydIF the release isn't botched (no pessimistic "minimum" VRAM requirements, emphasizing spooky shrinkwrap license with language about enabling others to bypass restrictions, or substituting a really inferior final upscaling stage instead of finishing training). It could unseat SD 1.5 as the default base model.
OTOH the samples I've seen so far are biased towards lame cinematic pablum. But I haven't tried it myself yet, so dunno the actual range of its aesthetics, and lots of people love that kind of stuff anyways.
SDXL is also significant in that it might be the last substantial/expensive-to-train locally-runnable image model to be released for a while - AFAIK there's nothing announced to be upcoming after this. Stability AI must be starting to run low on funds, and Runway and Sberbank were one-offs.
