LTX-Video: open-source real-time video diffusion, 5s of video in 2s on an H100
Crazy-Repeat-2006 · reddit · 2026-11-10
LTX-Video is a transformer-based real-time latent diffusion model for video, with source code and pretrained weights publicly available. Its core idea is to stop treating the Video-VAE and the denoising transformer as independent components and instead optimize their interaction.
Key design:
- A carefully designed Video-VAE achieves a 1:192 compression ratio, with spatiotemporal downscaling of 32×32×8 pixels per token — enabled by relocating patchifying from the transformer's input to the VAE's input
- In this highly compressed latent space, the transformer efficiently performs full spatiotemporal self-attention, essential for temporally consistent high-resolution video
- Since high compression limits fine details, the VAE decoder handles both latent-to-pixel conversion and the final denoising step, producing clean output directly in pixel space without the runtime cost of a separate upsampler
- Text-to-video and image-to-video are trained simultaneously
Performance: On an Nvidia H100, it generates 5 seconds of 768×512, 24fps video in just 2 seconds — faster than real time — outperforming all existing models of similar scale.
Blog: ltx.io/technical-blog; paper: arxiv.org/abs/2501.00103
Related event: LTX-Video Generates 5-Second Video in 2 Seconds on a Single H100(2 posts)→
More from Multimodal
- Agent clones 30s Kurzgesagt video in pure code: 720 frames, 99.4% frame match — RichardsonDx · 2026-11-10
- Fable 5.2 paints a photorealistic painting using roughly 700,000 simulated brush strokes — repligate · 2026-11-10
- fal launches FLUX 3 Max: tied #1 in image-to-video, 10s clips in ~12 seconds, 75% off launch — gorkem · 2026-11-10
- CreativeField: open-source ComfyUI interface unifying image, video and 3D workflows — QueasySide4011 · 2026-11-10
- NVIDIA DLSS 5 Wins CES 2027 Best of Innovation Award for 3D-Guided Neural Rendering — ctnzr · 2026-11-10
- Code-only motion graphics: 15-second scene made with Fable 5.2 for 4% of weekly quota — aziz4ai · 2026-11-10