LTX-2.5 Turbo 8-Step Distill LoRA

Lightricks' official rank-450 distill LoRA, downranked to run alongside the 22B dev transformer on a consumer GPU. It uses the same keys and delta as the original, in smaller files. Not an official Lightricks release.

How to use

1. Load the LTX-2.5 22B dev transformer (Lightricks/LTX-2.5) in ComfyUI. Never apply it to an already-distilled checkpoint.

2. Apply one file with LoraLoaderModelOnly (or LTX2LoraLoaderAdvanced) at strength 1.0. Try 0.8 if output looks over-sharpened.

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

Files

• ltx25_turbo_distill_r64.safetensors: ~1.32 GB. Smallest

• ltx25_turbo_distill_r128.safetensors: ~2.58 GB. Recommended start

• ltx25_turbo_distill_r256.safetensors: ~5.10 GB. Closest to the official file; use if you have the VRAM

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

Gallery: each video is one prompt and one seed at three ranks. Generated at 1MP and scaled to a 720 short edge, 121 frames at 24 fps. The audio plays r64 → r128 → r256 in turn.

Method: exact per-layer SVD of the official delta (B @ A). The rank budget is allocated adaptively ("water-filled"), so a file's label sets its total parameter count, not a cap on every layer. Small layers stay at r16–32, and some layers keep the full rank 450 in r256. The recon error of a uniform slice of the same size would be r64 62.3%, r128 50.0%, r256 33.0%. Files are bf16 and use the Comfy key layout (diffusion_model.transformer_blocks.0.ff.net.2.lora_A.weight), with 3320 tensors, the same as the official file.

Source: loras/ltx-2.5-22b-distilled-lora-450-bf16.safetensors in Lightricks/LTX-2.5. Related: pyros-vault/ltx-2.5-distilled-lora-450-reduced uses a different allocation and has not been compared side by side.

License: derivative of LTX-2.5, under the LTX-2 Community License Agreement. Entities with annual revenue of US$10M or more need a paid commercial license from Lightricks. Redistribution must include the license.