MiniMax H3, Part 26: Benchmark of 25 Acceleration Methods and New Tools

The 26th issue of the tool collection for the open video model MiniMax H3 in ComfyUI was released in the Telegram channel 'Neuronaut | Neural Networks in Creativity': the fast int8 video VAE by Kijai was updated, new LoRAs were added, and a benchmark of 25 acceleration methods on an RTX 5090 was published, with the best result being 1.71x for TaoMate 3-step distillation.

Compiled by Sergey KostenchukPublished 2026-09-21Updated 2026-09-22

2026-09-21 Multimodal HuggingFace

🎨 MiniMax H3, Part 26: Benchmark of 25 Accelerations and New Tools

A new issue of the collection on the open video model MiniMax H3 in ComfyUI: fast int8 video VAE by Kijai, Spectrum v0.2.28 with a CRLF fix on Windows, LoRAs ranging from 80s horror to character replacement from a single frame. A benchmark of 25 acceleration methods on an RTX 5090 was published: TaoMate 3-step provides 1.71x (32.7 seconds vs 56.2 seconds), sparse attention — 1.24x without quality loss.

🌍 The community around H3 — VAEs, quantization, distillation, caching — is growing faster than MiniMax's official releases: video speed on home GPUs is already a tunable parameter, not a vendor monopoly.

👤 For a local pipeline, this is a cheat sheet: acceleration up to 1.71x, the W4A8 version of ControlNet-Union compressed the checkpoint from 2.13 to 1.45 GB. Note: the measurements were taken on a single RTX 5090, and character replacement in someone else's video may affect rights.

Source 1: https://plox-1.github.io/h3-speedups Source 2: https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/vae

Sources