14/04: uploaded fp8mixed an nvfp4mixed of LTX-2.3-distilled-1.1
16/03: Lightricks released an official dev-nvfp4 and distilled-fp8
09/03: all formats of both dev & distilled in nvfp4, nvfp4mixed, fp8, and fp8mixed done.
07/03: ⚠️ make sure your torch (>=2.10) & cuda (>=13.0) are up to date, otherwise you won't see a speed boost using nvfp4, may even be slower ⚠️
05/03: Lightricks released an official dev-fp8, probably better to use it instead of mine.
Various formats of LTX-2.3, all created from official BF16 versions. Made with Silveroxides/convert_to_quant script:
fp8mixed
fp8
nvfp4mixed
nvfp4
non-calibrated / nc: direct conversion without SVD optimization
mixed: keep some weights in native format (embeddings, norm, bias, etc), same exclusion strategy as Lightricks
run time
config: Windows 11, Nvidia 5090, torch2.10+cu130, SageAttention2.2
- 8 steps, CFG 1
- default sampler/scheduler
- 960 x 544p, 121 frames, 24 fps
- x2 spatial upscale
- 4 refining steps
| format | distilled | dev + distill lora |
|---|---|---|
| nvfp4 | 54s | 57s |
| nvfp4-mixed | 57s | 59s |
| fp8 (mine) | 61s | 61s |
| fp8 (Lightricks) | TODO | 63s |
| bf16 | 87s | 91s |
Hot run, prompt already processed.
Smaller formats look a bit "fuzzy" around the edges in the outputs, but motion, audio, and lighting are still looking good.