modality cfg in ltx-2.3

[soccergame]

May I ask whether the modality CFG mentioned in the LTX-2 paper is supported?

I notice that there is a related implementation in the inference code of the official repository (https://github.com/Lightricks/LTX-2), but I could not find it in this repository.

It seems that None is being passed directly here (perturbations=None):

DiffSynth-Studio/diffsynth/models/ltx2_dit.py

Line 1682 in 84a75de

vx, ax = self._forward(video=video, audio=audio, perturbations=None, use_gradient_checkpointing=use_gradient_checkpointing, use_gradient_checkpointing_offload=use_gradient_checkpointing_offload)

Are there any plans to support modality CFG?

Thank you.

[Artiprocher]

@soccergame This is a novel, non-standard CFG implementation.

• During training, any CFG mechanism is inactive, so this does not affect the training process.
• During inference, the framework currently only supports the standard CFG implementation. If we were to support modality-specific CFG, we would need to verify its compatibility with other models in the framework, which requires time for validation. Conducting these experiments will take time and depends on our current workload.

[soccergame]

@Artiprocher Thanks for your reply and explanation. Looking forward to this support being added soon.