Buckets:
403 MB
3 files
Updated 23 days ago
Ctrl+K
| Name | Size | Uploaded | Xet hash |
|---|---|---|---|
| .gitattributes | 1.52 kB xet | 818ba6de | |
| README.md | 1.11 kB xet | 0f71ea14 | |
| lora_weights_step_04000.safetensors | 403 MB xet | 7d88f48e |
LTX-2 saxophone LoRA (rank 64, no-audio)
LoRA adapter for the Lightricks LTX-2.3 22B audio-video model, fine-tuned on a
saxophone-performance video dataset, with audio loss disabled during training
(with_audio: false) — this variant trains the video pathway only.
This is the exact checkpoint used to generate the v3/v4 with-LoRA side-by-side comparison clips for the no-audio ablation run.
Training
| Base model | ltx-2.3-22b-dev |
| Method | LoRA (rank 64, alpha 64) |
| Target modules | attn1/attn2: to_k, to_q, to_v, to_out.0 |
| Learning rate | 1e-3 |
| Step | 4000 |
| Effective batch | 8 (2×DDP × grad_accum 4) |
| Precision | bf16 |
| Mode | video-only (audio loss disabled) |
Files
lora_weights_step_04000.safetensors— the LoRA adapter weights.
Access
This repo is gated with manual approval — request access and the owner will review it.
- Total size
- 403 MB
- Files
- 3
- Last updated
- Jun 25
- Pre-warmed CDN
- US EU US EU