How much does the LTX-2.5 API cost?

The LTX-2.5 API bills per second of generated video: $0.09 per second at 720p, $0.15 at 1080p, $0.19 at 2K and $0.37 at 4K. Running the open weights yourself costs nothing in license fees for organizations under $10M in annual recurring revenue, so the API is for teams that would rather not manage GPUs.

How does LTX-2.5 compare to LTX-2.3?

LTX-2.5 replaces the VAE reconstruction step in LTX-2.3 with a diffusion video decoder, which lowers the measured visual-artifact score from 0.74 to 0.28 glitches per clip on Lightricks' own 98-prompt test. LTX-2.5 also adds native multi-shot scenes, automatic clip duration, a Gemma 4 12B text encoder, and native 4K HDR and RAW output.

Can I fine-tune LTX-2.5 on my own data?

Yes. LTX-2.5 ships a pretrained base checkpoint built specifically for fine-tuning, so teams can adapt it to their own domain and deploy it on their own infrastructure. Lightricks also released a checkpoint tuned for physical AI and robotics, aimed at teams whose training footage looks nothing like cinematic video.

What hardware does LTX-2.5 need?

LTX-2.5 needs a minimum of 16GB of VRAM and runs on ordinary GPUs, on-premises hardware, or edge machines. The headline 6.8-second figure for a 10-second 720p clip was measured self-hosted on two NVIDIA GB200 chips at steady state, so consumer cards are slower; the LTX API renders the same clip at 1080p in 23.7 seconds.

Where can I get LTX-2.5?

LTX-2.5 is available three ways: open weights on Hugging Face at Lightricks/LTX-2.5, native support inside ComfyUI from day one, and the hosted LTX API. The inference code lives in the Lightricks/LTX-2 GitHub repository, whose v1.2.0 release added checkpoint-driven architecture selection and NVFP4 quantization for Blackwell GPUs.

Lightricks · 2026-08-11 · major

LTX-2.5 — open-weights video model makes a 10s clip in 6.8 seconds

LTX-2.5 is a 22B open-weights video model that generates a 10-second clip in 6.8 seconds on two GB200 GPUs. It adds native multi-shot scenes, a new diffusion decoder, 4K HDR output, and runs from 16GB of VRAM.

Hugging Face model card banner for Lightricks LTX-2.5

An open-weights video model that renders connected multi-shot scenes faster than real time.

Key specs

10s clip, self hosted	6.8 s
Visual artifacts per clip	0.28

Quick facts

Maker	Lightricks (LTX)
Parameters	22B diffusion transformer
Text encoder	Custom Gemma 4 12B
License	LTX-2 Community License — free under $10M ARR
Min VRAM	16GB
Availability	Open weights on Hugging Face, native in ComfyUI, LTX API
What's new	Native multi-shot, new diffusion decoder, auto duration, 4K HDR and RAW

Pricing

API 720p	$0.09 / second of video
API 1080p	$0.15 / second of video
API 2K	$0.19 / second of video
API 4K	$0.37 / second of video
Self-hosted weights · Free under $10M ARR	$0

source ↗

What is it?

LTX-2.5 adds native multi-shot generation, so one prompt produces a sequence of connected shots that keep the same character, environment, lighting and voice across cuts instead of stitching separate clips together. The model is a 22B diffusion transformer with a custom Gemma 4 12B text encoder. Lightricks publishes the weights under the LTX-2 Community License, which is free for companies under $10M in annual recurring revenue.

How does it work?

A technique the company calls Diffusion Fidelity Rendering spends more compute on complex parts of a scene and less on simple ones. A new diffusion video decoder replaces the older VAE reconstruction step, which cuts glitches in high-motion shots and recovers fine detail like text and faces. An optional duration predictor picks the clip length that fits the requested action, and a retrained distilled variant reaches near-full-model quality with fewer retries.

Why does it matter?

Video generation has mostly been a closed API business, so teams could not fine-tune the model or keep footage on their own machines. Because LTX-2.5 ships its weights and needs 16GB of VRAM, a small studio can run the whole pipeline on hardware it already owns and adapt the base checkpoint to its own footage. Robotics teams get the same checkpoint for physical AI work, where the training data looks nothing like cinematic video.

Who is it for?

video teams, ComfyUI users, robotics and physical-AI researchers

Frequently asked questions

How much does the LTX-2.5 API cost?: The LTX-2.5 API bills per second of generated video: $0.09 per second at 720p, $0.15 at 1080p, $0.19 at 2K and $0.37 at 4K. Running the open weights yourself costs nothing in license fees for organizations under $10M in annual recurring revenue, so the API is for teams that would rather not manage GPUs.
How does LTX-2.5 compare to LTX-2.3?: LTX-2.5 replaces the VAE reconstruction step in LTX-2.3 with a diffusion video decoder, which lowers the measured visual-artifact score from 0.74 to 0.28 glitches per clip on Lightricks' own 98-prompt test. LTX-2.5 also adds native multi-shot scenes, automatic clip duration, a Gemma 4 12B text encoder, and native 4K HDR and RAW output.
Can I fine-tune LTX-2.5 on my own data?: Yes. LTX-2.5 ships a pretrained base checkpoint built specifically for fine-tuning, so teams can adapt it to their own domain and deploy it on their own infrastructure. Lightricks also released a checkpoint tuned for physical AI and robotics, aimed at teams whose training footage looks nothing like cinematic video.
What hardware does LTX-2.5 need?: LTX-2.5 needs a minimum of 16GB of VRAM and runs on ordinary GPUs, on-premises hardware, or edge machines. The headline 6.8-second figure for a 10-second 720p clip was measured self-hosted on two NVIDIA GB200 chips at steady state, so consumer cards are slower; the LTX API renders the same clip at 1080p in 23.7 seconds.
Where can I get LTX-2.5?: LTX-2.5 is available three ways: open weights on Hugging Face at Lightricks/LTX-2.5, native support inside ComfyUI from day one, and the hosted LTX API. The inference code lives in the Lightricks/LTX-2 GitHub repository, whose v1.2.0 release added checkpoint-driven architecture selection and NVFP4 quantization for Blackwell GPUs.

Try it

Download the weights from https://huggingface.co/Lightricks/LTX-2.5