SyFe LTX-2.3 ID-LoRA Checkpoints
Identity-driven audio-video LoRAs trained by SyFe on LTX-2.3 22B-dev. These checkpoints condition generation on a portrait and reference audio to preserve visual identity, vocal identity, and speaking motion in one pass.
Checkpoints
| Run | Training data | Resolution / frames | Rank | Steps | Status |
|---|
id_lora_ours_768 | 3,376 talking clips from one show | 768x320 / 121 | 128 | 3,000 | Experimental; cloned voice and mouth motion validated |
id_lora_ours_704 | 9,700 clips, 501 speakers, 20 shows | 1280x704 / 121 | 128 | 6,000 | Higher-resolution production candidate |
Each run folder contains the final LoRA and its training configuration. The adapters use the ID-LoRA audio_ref_only_ic contract with negative-time reference-audio conditioning. They require an LTX-2.3 ID-LoRA-compatible pipeline; they are not standalone models.
Limitations
Lip synchronization is functional but not phoneme-perfect, and speech may truncate when the requested line does not fit the video duration. Identity quality depends strongly on portrait framing and face size.
Use is subject to the LTX-2 community license and the applicable rights for all reference media and generated identities.
