ComfyStreamerH3 Runs Live Video Generation on a Single GPU

ComfyUI, the node-based AI workflow interface, now supports real-time video-to-video generation through a custom implementation called ComfyStreamerH3. Created by core developer pythongosssss, this workflow runs the MiniMax H3 video-to-video model locally on a single consumer GPU, achieving near-real-time playback speeds. The development represents a major shift for creators who previously relied on high-latency cloud APIs or frame-by-frame image models for live AI video generation.

What's new

The ComfyStreamerH3 project introduces a highly optimized pipeline designed to feed live video streams directly into ComfyUI. By utilizing Nvidia TensorRT and custom memory management, the workflow processes incoming video frames and applies the MiniMax H3 model with minimal latency. As of February 2025, tests demonstrate that a single Nvidia RTX 4090 GPU can sustain playable frame rates, bypassing the traditional batch-processing bottleneck that usually delays AI video generation by several seconds or minutes.

To achieve this speed, ComfyStreamerH3 uses a specialized streaming node that processes video in a continuous loop. Instead of waiting for an entire video clip to render, the system processes frames on the fly, applying temporal consistency directly to the live feed. This setup allows for immediate feedback loops, where changes to prompts or control parameters reflect in the video output almost instantly.

How it fits your workflow

For live streamers, VJs, and interactive installation artists, ComfyStreamerH3 offers a local alternative to expensive cloud-based APIs. Live video editors can now apply complex AI stylization to camera feeds or pre-recorded clips during live performances without worrying about network latency or per-minute API billing.

In terms of performance and stability, ComfyStreamerH3 serves as a direct alternative to running frame-by-frame Stable Diffusion workflows via StreamDiffusion. While older real-time ComfyUI setups struggled with flickering and temporal drift, the MiniMax H3 model natively understands motion, resulting in significantly smoother video output. This makes the workflow highly competitive against proprietary real-time video platforms like Fal.ai, giving creators complete local control over their data and rendering pipelines.

What it costs / how to try it

ComfyStreamerH3 is an open-source extension for ComfyUI and is entirely free to use. To run the workflow locally at acceptable frame rates, creators will need a modern Nvidia GPU with at least 16GB of VRAM (such as an RTX 3090 or RTX 4090) and must install the custom nodes along with the TensorRT dependencies via the ComfyUI Manager.