training script for the large_44k_v2 model

[ybx193670]

Can you provide the training script for the large_44k_v2 model, or what modifications need to be made to the current script in order to train the large_44k_v2 model?

[kszpxxzmc]

Also waiting here for a response

[hkchengrex]

Thank you for your interest.
Besides the modeling changes that are already in the code, we made the clip and the sync features null sampling independent. See

MMAudio/mmaudio/runner.py

Lines 265 to 268 in d1d78ad

	null_video = (samples < self.null_condition_probability)
	# complete mask is for when a video is not provided or we decided to ignore it
	clip_f[null_video] = self.network.module.empty_clip_feat
	sync_f[null_video] = self.network.module.empty_sync_feat

Note that the v2 model does not perform better in benchmarks. It just has a better "vibe" in generalization, according to me.