Minimax H3 Test: 27 seconds in 9 mins on local RTX 5090
Tokyo_Jab · reddit · 2026-08-31
Local test of Minimax H3 generating 27 seconds of video (0.8MP) in 9.3 minutes on an RTX 5090. Uses a two-stage latent workflow (4 steps at 0.4MP, 2 steps at 0.8MP) with Kijai's fast model, Turbo LoRA, and custom audio injection. Longer generations tend to fail or hallucinate.
More from Multimodal
- One Prompt, One Documentary: H3 Generates Hand-Drawn Video Explaining Espresso vs Americano vs Cappuccino — Best_Candidate_9060 · 2026-08-31
- Sori-1B: Audio-Grounded LM Trained From Scratch With No Text-Only Pretraining — Balance- · 2026-08-31
- Running MiniMax H3 on an 8GB laptop: full I2V pipeline with configs and failures — DG86 · 2026-08-31
- MiniMax H3 and H3 Max video models are 50% off on Vercel AI Gateway — evilrabbit_ · 2026-08-31
- MiniMax H3 Demo: Transferring Dancer to New Environment — CQDSN · 2026-08-31
- Infinite Livestream Built with Distilled FastH3 Model — boudaboy · 2026-08-31