Making the MiniMax H3 Video VAE 2x Faster
Blog post from Comfy
MiniMax H3’s video VAE has been updated to substantially reduce processing time on NVIDIA GPUs, achieving encoder speeds up to roughly 2.2 times faster and decoder speeds between 1.4 and 2.7 times faster, cutting a 1344×768, 129-frame encode-decode cycle from 24.3 to 12.7 seconds in testing on an RTX 5090. Improvements include a default fused encoder kernel that combines normalization, activation, and padding into one memory-efficient pass, custom convolutions that support the `--fast fp16_accumulation` option, and an int8 decoder that uses 8-bit weights and fused operations to limit intermediate memory transfers, including for attention. The changes are expected to benefit NVIDIA GPUs generally, while the fp16 accumulation option is especially useful on consumer cards with faster fp16 accumulation. Visual quality is described as effectively unchanged, with measured differences far below normal VAE reconstruction loss. Users can access the improvements by updating to ComfyUI version 0.36.0 or newer, enabling `--fast fp16_accumulation` for additional speed, and optionally replacing the standard VAE with the compatible int8 MiniMax-H3 VAE file.
No tracked trend matches for this post yet.
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.