RTX 3090 user cuts MiniMax H3 text-to-video to 7:39 with ComfyUI fixes and Sage attention
DeliciousGorilla · reddit · 2026-08-04
A user benchmarked MiniMax H3 on an RTX 3090 and cut a 10-second text-to-video run to 7:39 by fixing a ComfyUI import issue, using Sage attention through the KJ node, and switching away from nvfp4 to int8convrot on pre-Blackwell hardware.
Key findings
- comfy-kitchen 0.2.10 was silently breaking the ConvRot path with TensorCoreConvRotW4A4Layout import errors; upgrading to 0.2.26 fixed it.
- Disabling Sage attention was unnecessary on this setup; using it via the KJ node improved speed from 11.99 s/it to 9.29 s/it with no visible quality loss.
- The bundled nvfp4 text encoder has no hardware path on Ampere; for RTX 30-series cards, the author recommends int8convrot instead.
- The post includes a full workflow link and a Claude-generated prompt for the video.
More from coding & agent
- Photon connects Hermes Agent to iMessage in a 60-second setup — Teknium · 2026-08-04
- SkillHone keeps full decision history to push agent scores up on GAIA by 15.8 — jiqizhixin · 2026-08-04
- OpenComputer launches an early preview of its serverless agents runtime — zeeg · 2026-08-04
- Fable hooks into GitHub Stacked PRs and launches a 27-PR workflow — dbreunig · 2026-08-04
- A meme says the orchestrator in agent workflows is becoming the human — vikvang1 · 2026-08-04
- Users say Anthropic’s Fable 5 has regressed on harder coding tasks in the past week — _ghostchant · 2026-08-04