xlsd32-flow-beta1

opendiffusionai/xlsd32-flow-beta1

xlsd32-flow-beta1 is a text-to-image model from opendiffusionai. Use it when you need an image from a text prompt. It is set up for diffusers.

Similar to our sd-flow-alpha model, but hopefully better trained, AND this time it has SDXL vae bolted on.

Downloads · 30 days

7% of all-time downloads

All-time downloads

Public

Parameters

860M

5.5 GB on disk

Likes

Public

Repo makeup

Click a slice to open those files.

.safetensors5.5 GB · 100%

From the Hugging Face model README

Similar to our sd-flow-alpha model, but hopefully better trained, AND this time it has SDXL vae bolted on.

Started with our recent xlsd32-beta1 model which is already retrained for XL vae, so that part was out of the way.

Replaced noise scheduler with FlowMatchEulerDiscrete, then fine tuned, single run.

Around 160k images, "high res" (1536px+), square

B16a16, LR 5e-6, linear decay over 3 epochs (412 steps per epoch), using Optimi LION (sampled every 50 images) Chose the checkpoint at 1000 steps

I think thats pretty much it this time. No fancy stuff like prior version. Although I did use MY training code, at

https://github.com/ppbrown/ai-training/

but that shouldnt make much difference, I think?

Usable with diffusers pipeline. Will post demo code imminently, Use code from old version of this model if you are in a hurry