1⃣ MiniMax-H3 Video PixelHead v0.2.0: Lightweight experimental VAE adapter for video
2 min read
This model is a preliminary adapter (PixelHead) for the base video model Comfy-Org/MiniMax-H3, designed to accelerate generation and reduce VRAM consumption. Instead of the standard heavy 2.68 GB decoder, it offers a compact version weighing just 45.7 MB in BF16 format (or 91.4 MB in FP32). Currently, the build is an experimental preview: it handles flat graphics and general motion very well, but due to the lack of perceptual loss, it blurs fine details and natural textures. The development is intended for quick preliminary layouts and tests in ComfyUI.
Key features
- Compact size: The file size in BF16 format is only 45.7 MB compared to 2.68 GB for the stock decoder.
- Speedup: Provides about an 8.5x processing speed increase compared to the stock configuration.
- Flat object quality: Reliably decodes flat 2D graphics and dynamic text (metric reaches 36.73 dB).
- Integration: Comes with ready-to-use ComfyUI workflow files in English and Chinese.
What you can do
- Fast preview: Run a preview workflow in ComfyUI to evaluate motion and video composition when video memory is limited.
- Scene testing: Check generation drafts dominated by simple geometric shapes and static elements.
- Architecture exploration: Use the weights for experiments with local video generation pipelines on NVIDIA L4 cards.
Example
To run the preview version in ComfyUI, load one of the workflow files (for example, workflows/01_MiniMaxH3_VideoPixelHead_Preview_EN.json) and connect the downloaded weight file minimax_h3_video_pixel_head_bf16.safetensors to the decoder node.
Read more
DetailsBenchmarks (NVIDIA L4, 384x256 resolution, 22 frames): overall average PSNR is 31.03 dB, and average LPIPS is 0.2338.
Limitations: the current v0.2.0 version has issues with detail in complex textures, faces, and rapid occlusions (quality drops to 25.91 dB on such frames).
Hardware requirements: significantly reduced VRAM requirements due to the lightweight ConvNeXt backbone instead of the heavy ViT3D VAE.
Training details: the model was trained in pixel space using L1, MSE, and neighbor difference losses without perceptual LPIPS.














Comments