A 10-second video in 6.8 seconds — but read the test conditions

The speed chart measures image-to-video generation: time to produce a 10-second clip.

Three things need to be read alongside these numbers:

First, the 6.8-second figure ran on their cluster of two GB200s. The GB200 is Nvidia's latest data center GPU; one costs about as much as a car. It shows the model is efficient, but that's a different question from "how fast on your GPU"—and the official numbers don't cover that.

Second, the two LTX rows aren't the same resolution. The local row is 720p; the API row is 1080p. And that 23.7-second API time was measured end-to-end on third-party platform fal.run, including queue time.

Third, the competitors also had different conditions. Most ran at 720p, MiniMax H3 at 768p, Kling 3.0 Pro around 1080p, and the Veo 3.1 figure was for an 8-second clip, not a 10-second one.

Official chart: speed on the left, quality on the right. The small text under each row lists the test conditions for each vendor. Image / LTX

Audio and video are generated together

Most video models only handle the picture. Audio either doesn't exist or gets bolted on with a separate model. The LTX line has taken a different path since the previous generation: sound and image grow out of the same model at the same time. The company calls it a "video, audio, and world simulator."

The architecture has three layers:

The result isn't just dialogue and voiceover. It's a complete audio track that follows the characters, environment, style, and emotion, including ambient noise and foley. Other models shoot a silent film and add sound later; this one has the microphone on during the take.

The LTX-2.5 weights match this design: alongside the video VAE there's a separate audio VAE, and the capability tags include combinations like "text-to-audio," "video-to-audio," and "audio-to-video." The weight filenames carry 22b, confirming this generation's 22 billion total parameters. The previous generation was 14B + 5B; where the extra parameters went on the video side hasn't been disclosed.

The new rendering method: structure first, detail later

The main source of this generation's image quality improvement is called Diffusion Fidelity Rendering. It targets a chronic video generation problem: to run fast, the model doesn't work on raw pixels. It heavily compresses the video first, then generates; the more compression, the faster the math, but the blurrier it comes back, and faces are usually the first casualty.

LTX-2.5 splits "structure" and "detail" into two phases:

First in the quality chart — but it counts visible artifacts

The other chart measures visible artifacts per clip in text-to-video generation. Lower is better: LTX 2.5 Pro at 0.28, LTX 2.5 Fast at 0.39, Flux 3 at 0.45, MiniMax 0.46, Wan 2.6 at 0.65, Seedance 2.5 at 0.69, LTX 2.3 Pro at 0.74, Kling 3 Pro at 0.76, Veo 3.1 at 1.20.

LTX 2.5 Pro ranks first among the ten models. To see what that "first place" actually means, look at what was measured:

Three more methodological details are worth noting: 98 identical prompts were run across all 10 models; the scoring was done by a machine, not humans; and the chart is marked as preliminary, subject to change as the evaluation expands. Seedance 2.0's 0.60 score is also missing from the chart.

The three gates of "open source": revenue cap, fine-tune transfers, and handing over your email

The positioning is open weights, and the launch post puts it bluntly: "This is not a tool you rent. This is a foundation you build on. Open. Yours."

The actual terms have three layers:

None of these three conditions are particularly harsh; plenty of open models do the same. But outside the English-speaking world, "open source" is often read as "use it however you like"—and any one of these three gates could change whether this model fits your project.

On robotics, the official line is "still evolving"

The launch puts robotics and physical AI front and center: film, robotics, and real-time workflows are listed as three areas where LTX models are already in production. The CEO of robotics company Markov Robotics provides a testimonial, saying, "We haven't found another open model that can meet our needs the way this one does." The weights even include a separate pretrained checkpoint tuned for physical AI.

But the developer-facing documentation sounds far more cautious on the same topic:

Its established use is generating synchronized high-fidelity audio and video from text, image, and video inputs; applicability to emerging areas such as robotics and physical AI is still evolving.

LTX-2.5 Model Card · Hugging Face

Both sentences come from the same company. If you're building a robot, the second one is the more reliable measure.

On the film side, the company specifically names Asteria, an artist-led production house that built its hybrid workflow on LTX. Co-founder Paul Trillo says they value exactly the things that genuinely slot into existing production pipelines, like ACES. Image / LTX

How to get started: three paths, modular components, 16GB minimum

To run it yourself, there are three paths: ltx-pipelines (LTX's own PyTorch pipeline), ComfyUI (official workflow templates available on day one), and Diffusers.

The weights ship as a set of modular components rather than one monolith: the diffusion backbone, Gemma 4 text encoder, video VAE, audio VAE, a duration-control module, spatial and temporal upsamplers, and LoRA. During installation, each component needs its own path.

On hardware: 16GB VRAM minimum, any GPU works, and it supports local, edge, and API deployment. If you're short on VRAM, the workaround is dynamic precision reduction of the backbone plus CPU offloading. That 16GB figure's corresponding quantization level, resolution, and clip length aren't specified.

Quantized versions come in three flavors: an int8 build for ComfyUI, NVFP4 for Blackwell architecture, and runtime fp8 downcasting. The two ComfyUI-specific files won't work with the PyTorch pipeline, a note the model card calls out separately.

Within a day of release, the community had already posted 2 adapters, 4 fine-tunes, and 7 quantized versions on Hugging Face. Across the LTX family, cumulative downloads exceed 33 million (official figure, covering the whole series, not just 2.5).

The speed chart measures how long it takes to generate a 10-second clip. Per LTX's official self-evaluation, LTX-2.5 running locally on their cluster of two GB200s (Nvidia's latest data center GPU) took 6.8 seconds. On the same chart: Veo 3.1 at 70 seconds, Kling 3.0 Pro at 398 seconds.

This line has gone a different way since the last generation: a video stream and an audio stream run side by side, with continuous cross-attention (the model lets each stream reference what the other is doing), and both streams generate together on the same time schedule. The output isn't just dialogue — it's footsteps that match the visuals, ambient noise, and background sound that tracks the emotion.

Lightricks frames it as "a foundation you build on." The weights were indeed downloadable the same day, but the commercial terms are in the license, and they come in three layers.

The press materials put robotics front and center, listing film, robotics, and real-time workflows as three production use cases, with the CEO of robotics company Markov Robotics on record endorsing the model.