NYU Shanghai Library > RITS Home > *

LTX-2.5 Adds Native Multi-Shot Video and a Diffusion Decoder

August 12, 2026Provided by Utku Ege Tuluk

LTX released LTX-2.5 on August 11, 2026 — an update to its 22-billion-parameter open-weights audio-video model that replaces the VAE reconstruction stage with a diffusion video decoder, adds native multi-shot generation in a single pass, and ships with day-one ComfyUI support. The Israeli company behind Lightricks says the model generates a 10-second clip in 6.8 seconds on two NVIDIA GB200s — faster than real time — and it arrives with a pretrained checkpoint aimed at robotics fine-tuning.

Intermediate

A frame from an LTX-2.5 generated clip showing an enormous long-haired white yak with curved horns standing in an alpine meadow, with small human figures in blue and orange robes in the foreground and snow-capped mountains behind.
Image credit: LTX — sample output demonstrating Diffusion Fidelity Rendering

What Changed

The headline architectural change is the decoder. Where LTX-2.3 reconstructed frames from latents with a conventional VAE, LTX-2.5 substitutes a diffusion video decoder that LTX credits with “sharper faces, textures, and on-screen text, better motion, and fewer artifacts in demanding scenes.” The convolutional VAE is still shipped as a lighter alternative — the repository describes the diffusion variant as offering “improved quality at the cost of longer decode time and more VRAM.”

Extreme close-up frame of a human face generated by LTX-2.5, showing forehead and eyes in hard directional light with visible skin pore and hair detail.
Image credit: LTX — skin and hair detail from the new diffusion decoder

Sitting on top of that is what LTX calls Diffusion Fidelity Rendering, which allocates rendering compute according to scene complexity rather than spending it uniformly, operating in an 8× temporally compressed latent space. The text side changed too: LTX-2.3 required a separate Gemma 3 download, while LTX-2.5 bundles a custom Gemma 4 12B encoder (gemma4-12b-ltx-v1) plus an optional prompt enhancer that expands short prompts into detailed instructions, and a duration predictor that infers clip length from the described action before diffusion starts.

The most visible new capability is multi-shot. A single generation now produces several connected shots that hold character, environment, lighting, and voice across the cuts, rather than requiring separate generations stitched afterward.

Frame from an LTX-2.5 multi-shot sequence showing a figure in a hooded polar suit kneeling on ice beside an equipment case under a green aurora.
Image credit: LTX — multi-shot sample holding character and lighting across cuts

Specs and Vendor Benchmarks

The model ships as split, ComfyUI-aligned component files rather than the single bundles used in 2.3 — a 22B dev transformer for guided two-stage pipelines and a distilled variant that the quick-start runs at 8 steps in stage one and 4 in stage two at CFG 1. Frame counts must satisfy num_frames % 8 == 1 and dimensions must divide by 32; ComfyUI’s tutorial lists text-to-video, image-to-video, and first-last-frame-to-video workflows, with output supporting “native 4K HDR at up to 50 FPS.” int8 and NVFP4 quantizations, fp8 casting, and CPU offload are available for smaller cards, and LTX’s own comparison table claims a 16GB VRAM floor.

The performance numbers are LTX’s own and should be read as vendor-published. In the company’s image-to-video timing, a 10-second clip takes 6.8 seconds on-prem across 2× GB200 at steady state and 23.7 seconds through the LTX API, against 52 seconds for Omni Flash, 63 for Grok 1.5, 70 for Veo 3.1, 180 for MiniMax H3, 259 for FLUX 3, 317 for Seedance 2.5, and 398 for Kling 3.0 Pro. LTX notes the comparison is not like-for-like: competitor figures come from third-party host fal.run and include queue time, most render at 720p while the LTX API figure is an internal 1080p measurement, and the Veo 3.1 timing is for an 8-second clip. A separate artifact score over 98 text-to-video prompts, graded automatically and labelled preliminary, places LTX 2.5 Pro first at 0.28 and LTX 2.3 Pro seventh at 0.74 — the largest generational claim in the release is against its own predecessor.

Licensing and the Physical AI Angle

Frame generated by LTX-2.5 showing an orange industrial robot arm lifting a cardboard box from a warehouse shelving rack.
Image credit: LTX — sample from the physical-AI oriented pretrained checkpoint

“Open weights” here does not mean OSI-approved open source. LTX-2.5 is governed by the LTX-2.x Community License Agreement, effective August 11, 2026, which states that “entities with annual revenues of at least $10,000,000 (the ‘Commercial Entities’) are required to obtain a paid license,” with an exception permitting those entities to use the model for internal research and development. The agreement also carries modification-notice requirements and forbids removing or circumventing transparency features tied to synthetic-media disclosure obligations. Weights are on Hugging Face, in ComfyUI natively, and behind the LTX API; the company puts cumulative downloads across the LTX family at more than 33 million.

Alongside the standard checkpoints, LTX is publishing a pretrained foundation variant intended for fine-tuning on domain data — the target being physical AI and robotics rather than film. Co-founder and CEO Zeev Farbman framed the release as a deployment argument: “By keeping LTX open, we let teams own their hardware, their IP, and their model.” ComfyUI co-founder and CEO Yoland Yan, whose project carried the model on day one, said “Open is what lets the community move fast and lets businesses build on what it proves.”

What This Means

For researchers and students, the practical draw is the same one that made earlier LTX releases useful in a lab: a frontier-class video model whose weights you can download, quantize, and fine-tune on hardware you control, at a scale where a single well-specified workstation is enough to experiment. Multi-shot consistency in particular removes a step that previously had to be faked with careful seeding and manual continuity work, and the duration predictor and prompt enhancer shift some of the craft of prompting into the model itself.

The licence deserves attention before anyone builds on it. The $10M revenue threshold is generous for academic and small-team work, but it is a revenue-conditional grant rather than an open-source licence, and the distinction matters for anything that might outlive a research project. The benchmark spread also comes entirely from the vendor, with an acknowledged resolution and queue-time mismatch baked into the comparison — independent timings on commodity GPUs, rather than GB200 pairs, will be the more useful number for most readers.

Related Coverage

This post was drafted with AI assistance and reviewed by RITS staff.

Sources

Categories:
AI
Tags:
Level: Intermediate, ltx, open source, text-to-video, Video
Footer Logo

Connect with NYU Shanghai

New York University

Equal Opportunity and Non-Discrimination at NYU - New York University is committed to maintaining an environment that encourages and fosters respect for individual values and appropriate conduct among all persons. In all University spaces—physical and digital—programming, activities, and events are carried out in accordance with applicable law as well as University policy, which includes but is not limited to its Non-Discrimination and Anti-Harassment Policy.

沪ICP备15010846号-2beian沪公网安备31011502017015号

© 2025 All rights reserved.Powered by NYU Shanghai RITS