Update README.md
Browse files
README.md
CHANGED
|
@@ -16,4 +16,28 @@ Various formats for LTX-2.3, more to come.
|
|
| 16 |
|
| 17 |
- fp8: direct conversion from Lightricks weights
|
| 18 |
- nvfp4mixed: calibrated conversion, applying the same exclusion strategy as Lightricks
|
| 19 |
-
- nvfp4: calibrated conversion, targeting all transformer blocks, still some kept in native formats
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
- fp8: direct conversion from Lightricks weights
|
| 18 |
- nvfp4mixed: calibrated conversion, applying the same exclusion strategy as Lightricks
|
| 19 |
+
- nvfp4: calibrated conversion, targeting all transformer blocks, still some kept in native formats
|
| 20 |
+
|
| 21 |
+
---
|
| 22 |
+
|
| 23 |
+
## run time
|
| 24 |
+
|
| 25 |
+
config: Windows 11, Nvidia 5090, torch2.10+cu130, SageAttention2.2
|
| 26 |
+
|
| 27 |
+
- 8 steps, CFG 1
|
| 28 |
+
- default sampler/scheduler
|
| 29 |
+
- 960 x 544p, 121 frames, 24 fps
|
| 30 |
+
- x2 spatial upscale
|
| 31 |
+
- 4 refining steps
|
| 32 |
+
|
| 33 |
+
| format | distilled | dev + distill lora |
|
| 34 |
+
|------------------|-----------|--------------------|
|
| 35 |
+
| nvfp4 | 54s | 57s |
|
| 36 |
+
| nvfp4-mixed | TODO | 59s |
|
| 37 |
+
| fp8 (mine) | 61s | 61s |
|
| 38 |
+
| fp8 (Lightricks) | N/A | 63s |
|
| 39 |
+
| bf16 | 87s | 91s |
|
| 40 |
+
|
| 41 |
+
Hot run, prompt already processed.
|
| 42 |
+
|
| 43 |
+
Smaller formats look a bit "fuzzy" around the edges in the outputs, but motion, audio, and lighting are still looking good.
|