MiniMax H3 Video VAE Encodes Twice as Fast after Optimization
ComfyUI released an optimization update for the MiniMax H3 video VAE, doubling encoding performance and accelerating decoding on Nvidia hardware. Local AI video creators and workflow developers will experience much faster generation cycles when running high-resolution pipelines.
ComfyUI, the node-based AI workflow interface, optimized the MiniMax H3 video Variational Autoencoder (VAE) to run up to 2.2 times faster during encoding. This update addresses the heavy computational bottleneck of video compression and decompression, bringing local MiniMax H3 generations closer to real-time performance on Nvidia GPUs.
What's new
As of February 2025, the ComfyUI team implemented memory and compute optimizations specifically targeting the MiniMax H3 video VAE. The updated VAE now encodes video frames up to 2.2 times faster than previous implementations. Decoding performance also saw a major boost, running 1.4 to 2.7 times faster depending on the target resolution and Nvidia GPU architecture.
The optimization reduces the VRAM overhead during the latent-space translation phase. This means creators can process longer video clips or higher resolutions without hitting out-of-memory errors as quickly, making the local generation process much more stable on consumer-grade hardware.
How it fits your workflow
For filmmakers and animators building AI video generation pipelines in ComfyUI, VAE encoding and decoding are often the slowest parts of the render loop. Accelerating this step speeds up the iterative feedback loop when prompting and testing motion, allowing for faster prototyping of shots.
This optimization makes the MiniMax H3 workflow highly competitive with other open video ecosystems like Stable Video Diffusion (SVD) or LTX-Video. While platforms like Runway Gen-3 or Sora 2 Pro handle these computations on cloud servers, ComfyUI users running local hardware need every optimization possible to match cloud-based generation speeds. The faster VAE decoding directly reduces the time between the diffusion process finishing and the final video file rendering on screen.
What it costs / how to try it
ComfyUI is free and open-source. Users can access these speedups by updating their ComfyUI installation to the latest version via the ComfyUI Manager or running a git pull command. The MiniMax H3 model weights must be downloaded separately and placed in the appropriate models directory to use the optimized VAE nodes.
Read the original announcement on ComfyUI ↗