LTX-Video

LTX-Video (LTXV) is Lightricks' open-weights video generation family, released November 2024. It is the first DiT (diffusion transformer)-based video model engineered for faster-than-real-time generation: the flagship 2B model generates ~5 seconds of 24fps video at 768×512 in about 2 seconds on an H100 (provider-claimed), via a 1:192 spatiotemporal latent compression with a tightly integrated Video-VAE and denoising transformer (paper: arXiv 2501.00103).

Capabilities include text-to-video and image-to-video, keyframe conditioning, video extension (forward/backward), video-to-video, and distilled variants (13B and 2B distilled checkpoints in the 0.9.x line). The ecosystem includes ComfyUI nodes, diffusers integration, and quantized (Q8) builds that run on as little as ~6GB VRAM.

Benchmarks: Lightricks positions LTX-Video on speed and efficiency rather than standardized quality leaderboards — I found no provider or independent VBench-style scores for the 0.9.x line. The relevant "benchmarks" are throughput claims:

• ~5s of 24fps video in ~2s on an H100 (provider-reported, paper)
• Real-time 30fps generation at 1216×704 (provider-reported)

Note: LTX-2 (announced Oct 2025) is now Lightricks' primary model with synchronized audio+video and 4K output; LTX-Video 0.9.x remains available but is the legacy line.

No standardized (VBench-style) benchmark scores published by the provider or found independently. Performance claims are throughput-based: ~5s of 24fps video in ~2s on H100 (provider-reported, paper arXiv 2501.00103).