LTX-2.5, open video generation model runs on a single consumer RTX GPU

LTX said the model is optimized to run locally on NVIDIA RTX GPUs and NVIDIA DGX Spark. By lowering VRAM requirements, the company says it has made it possible to run a frontier-grade world model on hardware creators already own. The biggest change is "native multi-shot" generation. It renders a sequence spanning multiple shots as a single coherent output, keeping character appearance consistent from shot to shot. This addresses the flickering and breakage that made earlier open models hard to use for campaigns, according to the report. On top of that, a more refined Gemma 4 language backbone and a new decoder that reduces artifacts in fast motion push the output closer to a level usable without post-production, the report said.

The entire workflow runs inside ComfyUI, on a single consumer RTX GPU. Brand characters or signature styles can be locked in quickly through LoRA fine-tuning — a method that keeps the large model frozen and trains only small correction layers to fit a specific style. No studio, no cloud, and no IP ever leaves the device.

Speed makes the difference

For local operation to matter, speed has to back it up. According to an image-to-video benchmark LTX released, generation times for a 10-second clip are as follows.

Model/Method	Generation Time	Relative Speed
LTX-2.5 (on-premise, 2x GB200)	6.8s	2
LTX-2.5 (LTX API)	23.7s	6
Omni Flash / Grok 1.5 / Veo 3.1	52–70s	15
Seedance 2.0	196s	49
FLUX 3	259s	65
Seedance 2.5	317s	80
Kling 3.0 Pro	398s	100

On an on-premise basis, generation time is shorter than the clip's own runtime. That works out to about 7.6 times faster than the fastest closed alternative and roughly 58 times faster than the slowest. A speed gap of this size makes it genuinely feasible to batch out dozens of versions overnight and pick the best one the next morning. The report's core argument is that ad creative typically fatigues within 7 to 10 days — not for lack of ideas, but for lack of time and budget to produce enough of them.

The concept of a world model

While a large language model is trained to predict the next word, a world model is trained to predict the next moment. It generates an environment and simulates how objects and people move within it, letting users act inside that space. This structure is used not only in film, advertising, and games, but also in simulating robots moving through warehouses or factories. NVIDIA's open world model Cosmos 3, released on August 8, pointed in the same direction by combining vision reasoning with action prediction. LTX-2.5 differs slightly in that it compresses an actual production-scale video pipeline down to the scale of a single local GPU.

NVIDIA's August local AI series

LTX-2.5 is one pillar of NVIDIA's month-long "Local AI" series running throughout August. Released the same day, the open model Nemotron 3.5 Lightning is a 30B-parameter MoE (mixture-of-experts) model aimed at always-on agents, alongside the open-source library NeMo Switchyard, which routes each workflow step to the optimal model. The shared message is that open models are expanding hardware choice across RTX PCs, workstations, data centers, and the cloud.

What actually changes

The practical takeaway is that a solo creator or a small team can now match the output of a full-time production crew in a single day. With no per-clip charges or credit consumption, a GPU can churn out multiple versions overnight, and all that's left in the morning is to open a folder and pick one. Testing ten hooks and localizing for five markets is no longer a matter of booking a studio — it's a job for a single graphics card on a desk.