caroline bishop
January 22, 2026 14:48
NVIDIA releases a comprehensive guide to run FLUX.2 and LTX-2 visual AI models locally on RTX GPUs, reducing cloud costs and token fees for creators.
NVIDIA publishes detailed tutorials for running advanced visual generation AI locally on RTX PCs, positioning the hardware manufacturer as the go-to platform for creators looking to reduce cloud-based AI services and their associated costs.
Released on January 22, 2026, this guide walks users through setting up ComfyUI, an open source node-based workflow tool that is the backbone of local AI image and video generation. This timing coincides with recent RTX optimizations announced at CES earlier this month, which NVIDIA claims can significantly reduce generation times.
Why local power generation is important
The pitch is simple. Run your models on your own hardware, don’t pay per generation, and keep your assets directly under your control. For studios and agencies that have already integrated AI into their production workflows, eliminating “token anxiety,” the nagging concern about cost per use in the cloud, is a real operational benefit.
ComfyUI itself has been developing rapidly recently. Just last week, the platform added WAN 2.6 Reference-to-Video capabilities to mimic style and motion. Additionally, a FLUX.2 (klein) variant was released on January 15th with compact 4B and 9B parameter options optimized for faster local editing.
Hardware requirements and model selection
NVIDIA categorizes GPU requirements by use case. Key difference: RTX 50 series cards must use the FP4 model, while RTX 40 series performs best with the FP8 variant. This precision matching allows the model to consume less VRAM while maintaining performance.
FLUX.2-Dev can weigh over 30GB depending on the version, and large downloads require both storage space and patience. The tradeoff is photorealistic output quality, which NVIDIA says is now comparable to commercial cloud services.
For video generation, Lightricks’ LTX-2 model handles audio and video compositing with what NVIDIA describes as “storyboard-style” control. This model combines input images and text prompts, allowing users to take FLUX.2-generated images and animate them with specific camera movements, lighting, and even character dialogue.
Workflow benefits
ComfyUI’s node-based approach allows users to chain models together. This tutorial shows how to combine FLUX.2-Dev image generation directly into the LTX-2 video pipeline. This eliminates the manual steps of exporting the image, placing it on disk, and importing it into another workflow.
For users who have reached VRAM limits, ComfyUI now includes weight streaming, developed in collaboration with NVIDIA. This feature offloads parts of your workflow to system memory when GPU memory is full. Although performance will be lower, it is possible to generate it even on hardware that would otherwise stop working on large models.
NVIDIA also points to a 3D Guided Generation AI Blueprint for creators who want to incorporate 3D scene control into their image and video pipelines. This is a feature that brings local generation closer to production-ready tools.
A complete tutorial is available on NVIDIA’s blog, and additional prompt guides are linked from the Black Forest Labs and Lightricks documentation.
Image source: Shutterstock
