Nvidia shows how to run image and video AI models locally on RTX PCs
New guidance highlights how creators can generate images and video on their own computers, cutting cloud costs and speeding experimentation as more powerful graphics cards and models arrive.
NVIDIA has published instructions for running image and video generative AI locally on personal computers equipped with its RTX graphics cards, outlining a workflow that avoids reliance on cloud services.
In a statement, Nvidia said recent optimisations for RTX hardware, alongside new generative models introduced around the CES trade show, now allow creators to run advanced image and video workflows directly on their own machines. The company said this approach can reduce cloud computing costs and shorten the time it takes to iterate on creative ideas.
The guide focuses on using ComfyUI, a visual, node-based application that lets users build AI workflows by connecting components rather than writing code. Users are directed to download the software, open a starter workflow such as a basic text-to-image template, connect model nodes to an image output node and press “Run” to generate results. Prompts can then be adjusted to quickly try variations.
A key point for non-specialists is that ComfyUI itself does not include AI models. Instead, users download “model weights” separately. These weights are the billions of numerical values that encode what a trained model has learned. Because of their size, they can take up significant disk space. Nvidia noted that some versions of the FLUX.2 image model can exceed 30 gigabytes.
Once downloaded, the weight files are automatically saved to the correct folders for use by ComfyUI. From there, creators can experiment locally without sending prompts or images to external servers.
Nvidia also addressed the practical limits of running large models on consumer hardware. Different versions of models are optimised for different generations of RTX GPUs, and the company recommends lower-precision formats for newer cards to balance performance and memory use. Video models such as LTX-2 are particularly demanding, consuming large amounts of video memory.
To address that constraint, the guide points to a “weight streaming” feature that can offload parts of a workflow into a computer’s main system memory when the graphics card runs out of dedicated video memory. The trade-off, Nvidia said, is slower performance, but it allows complex workflows to run on hardware that might otherwise be unable to handle them.
The instructions also show how users can combine image and video models in a single workflow by copying and pasting nodes between saved projects. This allows, for example, an image generated with one model to be fed directly into a video generation step.
Related reading
- NVIDIA reveals blueprint for retail AI workflows
- NVIDIA's Drive automation debuts in Mercedes‑Benz CLA
- vLLM traces long-running memory leak to low-level system calls in disaggregated AI setup
Beyond the official guide, Nvidia directed users to community resources for troubleshooting and experimentation, including online forums and discussion channels where creators share workflows and tips.
The move reflects a broader shift in generative AI. While early tools relied heavily on cloud infrastructure, improved consumer GPUs and open-weight models are making it more practical to run sophisticated systems locally. For creators concerned about cost, speed or data privacy, Nvidia is positioning on-device generation as an increasingly viable alternative.
The Recap
- NVIDIA published guidance for running generative AI on RTX PCs.
- FLUX.2 model weights can exceed thirty gigabytes in size.
- Download ComfyUI then save workflows to generate images and videos.
