LTX-2.5

Highlights

• Video and audio generated jointly in the same model (not picture first, then dubbing)
• 22B DiT; official fp8 / nvfp4 and other VRAM-saving weights
• Templates cover T2V, I2V, FLF2V (first/last frame), and DFR multi-stage high resolution
• Official docs at docs.ltx.io include complete tutorials from getting started to multi-stage

Scope

• Good for: short shots with sound, product motion, atmospheric clips
• Good for: transitions controlled by first/last frames
• Not for: long videos (generation length per pass is limited)

Tips

• Set the first frame before the prompt; I2V is much more controllable than T2V
• Describe sound separately in the prompt (ambient sound/dialogue/music)
• On 16GB, do not enable DFR multi-stage; VRAM doubles

Cautions

• Community license; read LTX-2 LICENSE-2 before commercial use

Parameters

Resolution

Official templates provide tiers (mainly 720p class; DFR templates do high-resolution multi-stage)

Steps/CFG

Distilled version 8 steps, CFG 1; full version uses multiple steps (same rules as LTX-2.3)

Weight format

Full / distilled / fp8 / nvfp4

Deployment

16GB conclusion

The 22B original does not fit; fp8 or nvfp4 + block offload is required, and output is mainly 720p, a few seconds.

VRAM note

Running 22B video on 16GB means minutes per second of footage; start with the shortest duration to test parameters before extending.

Suggested starting point

First use the official I2V template + fp8 weights to produce a 3-second clip; after confirming the pipeline, then try DFR multi-stage.

Comparison

22B, joint audio-video, 2.03 million downloads. The division of labor with MiniMax H3 is clear: H3 is strong in multi-reference and audio detail, while LTX-2.5 is strong in complete templates (four sets: I2V/T2V/FLF2V/DFR) and fp8/nvfp4 VRAM-saving weights, making it easier to run on 16GB.