search

axodox / axodox-machinelearning

This repository contains a pure C++ ONNX implementation of multiple offline AI models, such as StableDiffusion (1.5 and XL), ControlNet, Midas, HED and OpenPose.
MIT License
605 stars 35 forks source link

Support v-prediction objective, fix some SDXL models outputting clouds #26

Open ZDisket opened 7 months ago

ZDisket commented 7 months ago

If you try to use models trained with the v-prediction objective with the current repo, you'll get nonsense output. I detailed in this issue This PR adds v-prediction to both schedulers, and aligns the Euler Ancestral one closer to the HF implementation (the original doesn't use predictedOriginalSample).

Also, with some SDXL models, it will output nonsense/clouds. This is Pony Diffusion V6 XL 042-sc256

This is because some of them demand that we sample from the penultimate layer on both text encoders, which the Huggingface implementation does by default. This change fixes that model and (feels like) it makes other SDXL models follow prompts better.

axodox commented 6 months ago

Thanks, I will look into it.

axodox commented 6 months ago

I am preparing a larger reorganization of the code on a branch to make it more flexible for future development, I will migrate your change once I finish it.

Currently I am also working on a change for DirectML ONNX executor which should reduce VRAM usage a lot, particularly for VAE decode, where it could be reduced by 75% or more.