NVIDIA and ComfyUI have partnered to bring new most wanted features to ComfyUI via upscaler nodes and nvfp4 versions of popular image generation models. This collaboration improves how artists and developers work with generative AI for image and video production. ComfyUI is a powerful, open-source tool that uses nodes to create workflows. The following updates simplify this process and increase performance.
Simplified User Interface
Artists can now use a new feature called App View. This interface presents workflows in a standard way. Users only need to enter a prompt, adjust simple parameters, and generate an image. The traditional Node View remains available for advanced users. You can switch between these two modes seamlessly. App View is available today.
RTX Video Node
ComfyUI now includes an RTX Video node. This tool allows you to generate video more efficiently. Instead of creating a time consuming 4K video directly, you can generate a lower resolution version first and then upscale the resolution to 4k in seconds. You can now iterate on videos as fast as you iterate on images.
Download NVIDIA 4k upscaler package
NVIDIA released a free Python 4k upscaler package for developers. You can access this via the PyPI repository. Sample code is available on GitHub, along with a VFX Python bindings guide. These resources help you get started quickly. The package provides programmatic access to the same AI upscaling technology that powers RTX Video. It runs directly on RTX GPU Tensor Cores of blackwell(fast) and previous gen tensors a little bit slower. This delivers 4K upscaling 30 times faster than alternative local upscalers. It also costs a fraction of the VRAM required by other tools. Download link available as MIT license, repo link: https://pypi.org/project/nvidia-vfx/
Download native nvfp4 local image generation models 2026:
NVIDIA has collaborated with ComfyUI to enable NVFP4 and FP8 support directly within the app. This technology compresses model weights. This change accelerates performance by over 2X. It reduces memory consumption by 60%. The quality does not significantly degrade. You can now run massive, high-fidelity models on standard RTX GPUs. These GPUs previously could not handle the memory requirements. This enables professional-grade generation directly at your fingertips.
Getting Started with New Models
More NVFP4 and FP8 models will be released today. These include LTX-2.3(coming shortly at that link), FLUX.2 Klein 4B, and FLUX.2 Klein 9B. To use these models, follow these steps:
- Download the NVFP4 and FP8 checkpoints from Hugging Face.
- Open ComfyUI and access the Template Browser.
- Load the default workflows.
- Replace the default model checkpoint with the new file you downloaded.
