Is v0.9.5 backward compatible?

[Luvata]

Congrats on v0.9.5 another banger!

Could you point me to the technical changelog for the new LTX-Video v0.9.5? Beyond the improved VAE and higher resolution already mentioned in the README, since it seems to support many more capabilities and controllable features.

Also, is it backward compatible with previous checkpoints? Specifically, can I directly load the new checkpoint using the previous implementation (e.g., diffusers), assuming I don’t need the prompt enhancement?

[suhyeok-jang]

totally agree.

[Luvata]

I tried loading v0.9.5 directly with diffusers 0.32.2
pipe.transformer = LTXVideoTransformer3DModel.from_single_file("https://huggingface.co/Lightricks/LTX-Video/ltx-video-2b-v0.9.5.safetensors", torch_dtype=torch.bfloat16)
It can still return a video file, but it’s very noisy, so there seem to be some updates to the model

[suhyeok-jang]

any updates?
@yoavhacohen

[nitinmukesh]

It is backward compatible. There is also a new pipeline.
huggingface/diffusers#11104

[yoavhacohen]

inference.py is fully backward-compatible with previous checkpoints.

All checkpoints (0.9.0, 0.9.1, and 0.9.5) use the same transformer architecture, but their VAEs differ.

The PR updating Diffusers to support LTX-Video 0.9.5 was merged today:
huggingface/diffusers#10968