LORA Trainable?
Are we able to train a LORA from samples using your model? Do you have any advice on how do to that?
We haven't released training code yet, but LoRA fine-tuning on this architecture is definitely possible. Resemble AI's DramaBox project does exactly this on the same LTX 2.3 audio base with
IC-LoRA, so there's a reference implementation out there if you want to get started now.
The key things to keep in mind: target the self-attention layers (not cross-attention, you want to preserve prompt adherence), and use rank 128. If you're training for a specific voice identity,
include 10-20 seconds of reference audio in ~30% of your training samples.
We haven't released training code yet, but LoRA fine-tuning on this architecture is definitely possible. Resemble AI's DramaBox project does exactly this on the same LTX 2.3 audio base with
IC-LoRA, so there's a reference implementation out there if you want to get started now.The key things to keep in mind: target the self-attention layers (not cross-attention, you want to preserve prompt adherence), and use rank 128. If you're training for a specific voice identity,
include 10-20 seconds of reference audio in ~30% of your training samples.
Are you guys planning to release the training code? I want to fine-tune it with Hindi, as it lacks it.