Local Minimax H3 2K video generation on a 5090: 1984x1120 in 30 minutes, full prompt shared
Kooky-Mode3047 · reddit · 2026-10-06
A Reddit user shares a hands-on local run of Minimax H3 video generation on an RTX 5090 at 2K (1984x1120, 2.09mpx): 25 steps, Euler+Beta, 30 minutes for a 13-second scene.
- Tradeoffs: Euler+Beta was chosen over seeds2 + sgmuniform (better prompt following but 1 hour); the author accidentally used the 20-49 hybrid fl2va/ref2va instead of 30-49, leaving some visual quality on the table.
- Spectrum acceleration: guessing about half the steps greatly speeds things up but hurts prompt following; with seeds2, Spectrum forecasts H3's internal calls (2N-1), giving strong results at double the runtime.
- Structured prompt: the full YAML-style prompt is shared, with subjectdefinitions, summary, retentionanalysis, and detaileddescription sections locking character identity and attire from reference images (a Castle-style scene) and tracking preservation shot by shot; prompts had to be massaged to compensate for H3's shaky denoising trajectory with Euler.
- Full 1080p video is on YouTube.
More from Multimodal
- Midjourney tilt-shift prompt turns a morning kitchen into a toy diorama — tisch_eins · 2026-10-06
- huashu-art-motion: A Claude Code Skill That Produced a 23-Style Animation in Under a Day — AlchainHust · 2026-10-06
- inclusionAI open-sources Ming-Image-0.1-Design: 6B text-to-image model with transparent RGBA output — lmoroney · 2026-10-06
- Wan2GP memory management update: 15s video now renders in 5 minutes, first-click OOM fixed — orangpelupa · 2026-10-06
- MIRRORSIDE: a behind-the-scenes film set that never existed, made with AI — StrategyMedium5907 · 2026-10-06
- Qwen-Image 2.1 edits look unfinished: user shares ComfyUI params seeking fixes — Suspicious_Aide2697 · 2026-10-06