ComfyUI makes the MiniMax H3 video VAE up to 2.2x faster on Nvidia GPUs
NerdyRodent · x · 2026-11-17
ComfyUI's update speeds up the MiniMax H3 video VAE to 2.2x faster encoding and 1.4-2.7x faster decoding on Nvidia GPUs, halving VAE time for a 1344x768, 129-frame round trip (24.3s → 12.7s). Changes include a fused encoder kernel (1.5x), custom convolutions honoring --fast fp16accumulation with folded bias/skip connections (→2.2x), and an int8 decoder with 8-bit attention (1.4x over int8). Quality loss is imperceptible.
More from Infra
- Qdrant Generates 32.8M Embeddings in 35 Minutes With Static Embedding Models — qdrant_engine · 2026-11-18
- ArisGlobal cut Bedrock inference cost bloat by caching shared prompt prefixes first — AWS ML Blog · 2026-11-17
- Cloudflare researcher: Clef's fast release was 'asymmetric effort,' reasoning models next — michellechen · 2026-11-17
- Internet meltdown day 4: brokers hit as Charles Schwab outage blocks traders — GaryMarcus · 2026-11-17
- Chinese CSPs slash 2027 enterprise SSD orders, NAND flips to oversupply with prices falling from 2Q27 — SumitGup · 2026-11-17
- Running BGE-M3 + reranker + TTS in 4GB VRAM and ~300MB RAM — ALexiosK11 · 2026-11-17