LTX-2.5 Turbo 8-Step Distill LoRA

Official rank-450 distill, downranked so it sits next to the 22B dev transformer on a consumer GPU. Same keys. Same B @ A delta. Smaller files.

8 steps. CFG 1. Apply on dev, never on an already-distilled checkpoint. Not an official Lightricks release.

Rank gate

Each video example is one prompt, one seed, three ranks. Generated at 1MP (multiples of 32), then scaled so the short edge is 720. Length is 121 frames at 24 fps (~5.04s), looped three times.

[ prompt + settings ][ Rank 256 ]

[ Rank 128 ][ Rank 64 ]

The soundtrack is concatenated r64 → r128 → r256. The speaker lamp on the cell is the rank you are hearing.

Start with r128. Use r256 when VRAM allows. r64 is the small pack.

Files

ltx25_turbo_distill_r64.safetensors~1.32 GB

ltx25_turbo_distill_r128.safetensors~2.58 GB

ltx25_turbo_distill_r256.safetensors~5.10 GB

Source: Lightricks’ official r450 distill LoRA in Lightricks/LTX-2.5 (loras/ltx-2.5-22b-distilled-lora-450-bf16.safetensors, ~8.9 GB).

Each pack has 3320 tensors (1660 modules × A/B), matching the official file. A uniform slice at the same size is worse: r64 62.3%, r128 50.0%, r256 33.0% recon error. Adaptive water-fill is the quality-per-byte path.

How the downrank works

1. Source. The official r450 LoRA already is the distill delta.

2. Exact SVD per layer. For each module, W = B @ A is factored without forming the full matrix: QR both sides, SVD the small core. Optimal rank-k approximation of that layer’s official delta.

3. Water-fill the rank budget. A file labeled r64 / r128 / r256 has the same parameter count as a uniform rank-64 / 128 / 256 slice — not a hard cap of that rank on every layer. Tiny layers stay at r16–r32. Remaining budget goes where the next singular value buys the most energy per parameter. At the r256 budget some layers keep the full official 450.

4. Save bf16 lora_A / lora_B. Same Comfy key layout. Strength 1.0.

How to use

1. Load the LTX-2.5 22B dev transformer from Lightricks/LTX-2.5 in ComfyUI.

2. Apply one file above with LoraLoaderModelOnly (or LTX2LoraLoaderAdvanced), strength 1.0 (try 0.8 if it looks over-sharp).

3. Sample like distilled: 8 steps, CFG 1 (video and audio).

For maximum fidelity use native distilled or the official r450 LoRA.

Key format

diffusion_model.transformer_blocks.0.ff.net.2.lora_A.weight
diffusion_model.transformer_blocks.0.ff.net.2.lora_B.weight

Caveats

• Downrank, not identity. Rank 450 is high-rank; r64/r128 still throw away a lot of the official delta. r256 is the closest pack here.

• Scheduler matters. Evaluate at 8 steps / CFG 1.

• Related public reduce: pyros-vault/ltx-2.5-distilled-lora-450-reduced (energy-threshold + rank ceiling). Different allocation; no claim we beat it without a side-by-side.

License

These adapters are Derivatives of LTX-2.5 (including the official distill LoRA) and inherit the LTX-2 Community License Agreement. Entities with annual revenue ≥ US $10,000,000 need a paid commercial license from Lightricks before using LTX-2.x or derivatives. Redistribution must include this license.

Citation

@article{hacohen2025ltx2,
  title={LTX-2: Efficient Joint Audio-Visual Foundation Model},
  author={HaCohen, Yoav and Brazowski, Benny and Chiprut, Nisan and others},
  journal={arXiv preprint arXiv:2601.03233},
  year={2026}
}

Base model and official r450 LoRA: Lightricks/LTX-2.5