[
LTX-2.5 Model
]
LTX-2.5 improves quality, continuity, control, and efficiency through native multi-shot, stronger prompt adherence, and better local performance.
LTX-2.5 Model
LTX-2.5 improves quality, continuity, control, and efficiency through native multi-shot, stronger prompt adherence, and better local performance.
Introducing
A new video generation technology that allocates rendering compute by scene complexity, delivering industry-leading pixel quality that holds up frame by frame, even on a cinema screen.
Better control, stronger continuity, and more usable results.
Fewer motion artifacts. Smoother, more natural results.
Create connected shots that hold character, environment, lighting, and voice.
Follow complex creative instructions from shorter, simpler prompts.
Generate the right clip length for the requested action.
Generate high-resolution HDR footage built for professional finishing.
Generate and edit inside professional color and finishing pipelines without compromising the master.
Get near-full-model quality with fewer retries, fewer controls, and significantly lower compute.
Fine-tune a raw foundation model on your own data and domain.
What changed: New diffusion video decoder.
What changed: Gemma 4 12B text encoder + custom prompt enhancer.
What changed: Automatic duration prediction.
What changed: Improved distilled model, larger dataset + RL, prompt enhancement, and automatic duration prediction.
Comparison
Run it on your hardware. Fine-tune it on your data. Edit and extend it on your terms.
LTX
MiniMax H3
Seedance 2.5
Kling 3.0 Pro
Veo 3.1
Grok 1.5
Image-to-video, time to generate a 10s clip. Lower is better.
On-prem runs on our own hardware, 2x GB200, at steady state. API figures are measured end to end on fal.run, a third party provider, and include queue time. Image to video, 10s target, 24fps. Resolutions vary. Omni Flash, Grok 1.5, Veo 3.1, Seedance 2.0, and FLUX 3 render at 720p. Seedance 2.5 tops out at 720p. MiniMax H3 renders at 768p with no true 720p tier. Kling 3.0 Pro renders at roughly 1080p. Veo 3.1 is measured on an 8s clip rather than 10s. LTX API figure is an internal measurement at 1080p. It renders at a higher resolution than most models here and is still the fastest.
Text-to-video, visible glitches per clip. Lower is better.
The artifact score counts blotches, broken texture, and melted or smeared areas. The same 98 prompts, text to video, run through 10 models and graded by automated scoring rather than human viewers. LTX 2.5 Pro ranks first of the ten. Seedance 2.0, at 0.60, is not shown. Preliminary results, expected to evolve as evaluation expands.
Try the model, build it into your product, or run the full stack yourself.
Run locally, fine-tune, and deploy on hardware you control. Free under $10M ARR.
Open weights, open workflows. Built and shared by the LTX community.
FAQs
LTX-2.5 is the newest version of LTX, an open-weights world model for video generation. It's built as a stronger foundation, not just a new tool, that teams can build on, fine-tune, and deploy themselves.
It adds a better video decoder for cleaner motion, native multi-shot scenes that stay consistent across cuts, stronger prompt understanding, automatic clip duration, beta precise video editing, native 4K HDR and RAW support, and a faster, higher-quality distilled model.
Yes. LTX-2.5 includes a pretrained base checkpoint built specifically for fine-tuning, so you can adapt it to your own domain and deploy it on your own infrastructure.
Self-hosted on 2× GB200 GPUs at 720p, LTX-2.5 generates a 10-second video clip in 6.8 seconds, which is faster than real time.
API pricing is billed per second of generated video. ltx-2-5-fast: $0.09/sec at 720p, $0.13/sec at 1080p, $0.19/sec at 1440p, and $0.30/sec at 4K. ltx-2-5-pro: $0.12/sec at 720p and $0.17/sec at 1080p. Full rates are on the API pricing page.
LTX-2.5 uses open weights with a permissive license, no mandatory branding, and free use under $10M ARR.
