This LoRA works well with LTX 2.3 I2V, I haven't tested it with T2V. It is audio only, so you have to ask for the sound effect by using the trigger word "".

Training specs:

• Mode: T2A (audio-only) — targets only audio_attn1, audio_attn2, audio_ff modules, doesn't touch video or speech-related weights
• Rank 16 / alpha 16, trained unquantized (bf16) for 2000 steps
• Dataset: 29 training clips (of 32 total — 1 too short for any bucket, 2 held out for validation) that I generated and hand-captioned with texture descriptions (e.g. "long wet," "small," "walking triple") rather than the literal word "," so the trigger word alone carries the concept
• Final loss 0.5122, trained in ~8.4 hours overnight
• Validated against 8 prompts (3 in-distribution, 3 novel out-of-distribution, 2 held-out)