v0.11.3

LTX is about 2.7x faster than 0.11.2 with the new default model, and 1.75x faster even if you stay on High Quality model (old default).

Support on Buy Me a Coffee or Unifans. More info
中国用户请通过闲鱼支持。其他渠道都是骗子。

Thanks to all the supporters for making it possible to train two new distilled models so quickly!

感谢所有支持者，让我们得以如此迅速地训练出两个新的蒸馏模型！

New default: Standard Quality model

Two new LTX models. Standard Quality is the new default and the fastest. Quality looks very close to me, do A/B tests and tell me if you see a difference.

Model	Speed vs Standard
Standard Quality	1x	new, default
Good Quality	~1.5x slower	new
High Quality	~2.5x slower	old default (it's still 1.75x faster compared to v0.11.2)
Top Quality	~5x slower	old Best

Speeds are for long videos. On short videos the gap is smaller.

How much faster

Same 40 s 1080p video on RTX 5090, release builds:

	0.11.2 (High)	0.11.3 (High)	0.11.3 (Standard, new default)
Whole job	186 s	106 s	68 s
Speed-up		1.75x	2.7x

Most of it comes from step 2 and step 4, which every model uses:

GPU	Step 2, per window	Step 4, per window	Step 4, Sharper large mosaics
RTX 5090	0.44 → 0.36 s	1.17 → 0.73 s	2.98 → 1.64 s
RTX 4090	0.64 → 0.49 s	1.84 → 1.07 s	4.35 → 2.45 s
RTX 5080	0.89 → 0.73 s	2.66 → 1.50 s	6.31 → 4.65 s
RTX 3090	1.37 → 1.19 s	3.72 → 2.26 s	9.10 → 5.09 s
RTX 4060 8 GB	3.04 → 2.48 s	8.83 → 6.45 s	-

Step 4 also writes the video while the next window decodes, on the 5090 the whole step went from 41 s to 19 s on that video.

RTX 50xx: Standard and Good also come in Fast mode (NVFP4), like High and Top. Fast mode is about 20% faster per step and uses about 5 GB less VRAM, so it helps most on long videos.

Other changes

• progress bar no longer jumps back and time left is more accurate
• asks before stopping, and offers to minimize when you close the window during processing
• LTX fixes for 6-8GB GPUs and long videos
• VC-1 videos on NVIDIA no longer drop frames
• videos with broken metadata text open again
• optimize end of phase 1 for LTX (could appear as stuck at 99%)
• streaming fixes
• fixed a freeze when stopping or cutting videos that are decoded on the CPU (AMD and some formats)
• Windows: activation and startup fixes, right font for Chinese text in the log
• Linux: ZFS support
• Optimize release size (linux only)
• Optimize model load time (matters most on shorter clips where loading could take most of the total processing time)

Updating

• The first run after updating is slower, it prepares the GPU once per LTX model variant.
• You can copy the LTX models from your old model_weights folder to skip the big downloads. Standard is a new model, so it downloads once.
• CLI: ltx is Standard now, ltx-4step is Good, ltx-8step is High (old default), ltx-undistilled is Top. Saved settings keep the model you picked.