pyros-vault/ltx-2.5-distilled-lora-450-reduced

LTX-2.5 distilled LoRA-450 — rank-reduced (ceil72 / ceil128)

Rank-reduced versions of the official Lightricks/LTX-2.5 loras/ltx-2.5-22b-distilled-lora-450-bf16.safetensors, for few-step ("turbo") sampling on consumer GPUs where the full 8.9 GB rank-450 file doesn't fit next to the model during inference or training previews.

File	Size	Rank ceiling	Median kept energy
..._fro90_ceil72.safetensors	1.38 GB	72	0.71
..._fro90_ceil128.safetensors	2.31 GB	128	0.82

Start with ceil72 (validated in training previews at weight 0.8 with the official 8-step sigma schedule); use ceil128 if you want closer fidelity and have the headroom.

Recipe

Per module, the delta B @ A is truncated by factored SVD (QR of both factors + core SVD — mathematically optimal rank-k approximation): singular values kept until 90% Frobenius energy, capped at rank 72, with a small condition safeguard. Same key format as the original (diffusion_model.*.lora_A/lora_B), drop-in wherever the original loads.

Honest quality numbers

The rank-450 distillation delta is genuinely high-rank; the ceiling binds for most modules. Per-module retained Frobenius energy (1,660 modules):

	ceil72	ceil128
ceiling-capped	74%	63%
median kept energy	0.71	0.82
5th percentile	0.46	0.61
worst module	0.38	0.54

Full per-module breakdowns in the reduce_report_*.json files. Despite the haircut, distillation behavior is redundant across directions in practice — the ceil72 file renders clean 8-step previews. A spectral check against the official distilled transformer confirms the rank-450 LoRA is the exact source of that checkpoint (the delta spectrum cliffs at rank 450), so SVD truncation of this LoRA is the optimal reduction — there is no extra signal to extract from the full model. If you need maximum fidelity, use the official full-rank file. Typical usage: weight ~0.8 with the official 8-step sigma schedule.

Provenance

Derivative of the official Lightricks release; the LTX-2.x Community License applies. Produced with ltx2_reduce_lora.py (musubi-tuner LTX-2 branch).