Show HN: Unstock.ai – Free image generation with SDXL

by mmoustafa • 3 years ago

9 6 3 years ago

Hi HN!

I've got the barebones of a service running on top of Stable Diffusion XL. I can cheaply run image generations at 1024x1024. And of course there's a limit to how fast I can generate them given the request queue and limited GPUs, but the service is cheap enough that I'm happy to run it out of pocket for now.

Let me know your thoughts, I hope you enjoy the service!

ilaksh• 3 years ago

The web page says DALL-E still. Are you using the Stability.AI API or ClipDrop or RunPod or Replicate or what? Just curious.

mmoustafa• 3 years ago

Sorry that was a typo when I started out, for now I'm running on a couple different GCP GPU instances

itake2• 3 years ago

I tested SDXL this weekend with stable-diffusion-webui and I wasn't impressed. There are multiple SD1.5 checkpoints that were much better at generating pics of humans.

Is this issue that SDXL needs more fine tuning for it to be useful and it will be better than SD1.5 once it is?

dragonwriter2• 3 years ago

Is this issue that SDXL needs more fine tuning for it to be useful and it will be better than SD1.5 once it is?

No, its that SDXL is better than SD1.5 right now, but some SD1.5-derived-checkpoints are still better than base SDXL, at least foe their focal use cases. There are already a bunch of SDXL-derived checkpoints, but obviously not as much effort yet into them as SD1.5 (though a lot more being pumped out than ever happened for SD2.x).

(Also, while I think some of the forks do, base stable-diffusion-web-ui doesn’t have a built-in workflow that uses the SDXL refiner model, to get full SDXL effect you have to generate with the base model and separately do img2img with the refiner model.)

itake2• 3 years ago

Are there any SDXL opensource codebase (or self-hosted apis) that support LORA models?

Show HN: Unstock.ai – Free image generation with SDXL

Related Stories