Minimax H3 workflow benchmarks: Seed Hunter path hits 2.14x speedup over 20-step baseline
BluePointDigital · reddit · 2026-09-02
The author ran 12+ hours of agent-driven workflow comparisons on Minimax H3 video generation and published the full dataset: 16 curated video-and-metric cards, 151 sanitized timing records, the canonical prompt, methodology notes, and machine-readable JSON/CSV on GitHub Pages.
The main test uses a deliberately hard 15.08s prompt (768×1344, 24fps, 362 frames) combining a talking selfie shot, exact dialogue, walking motion, a rapid camera pan, a multi-subject vehicle collision, and a return to the speaker—designed to expose identity drift, bad anatomy, motion breakdown, camera-continuity issues, dialogue changes, and lip-sync problems.
Timing vs the 20-step baseline:
- SageAttention2 + FirstBlockCache Safe (20 steps): 10:11.4, 1.00x
- PDD + Sage (8 steps): 6:15.0 median, 1.63x
- Seed Hunter direct one-seed path (12+4 steps): 4:45.8, 2.14x
The author stresses these are local measurements, not universal claims—and fastest isn't best: one 4:03.5 path developed a visible perspective error during the crash, while a camera-POV prompt clarification produced a far more coherent 4:25.3 result. Reproducible H3 results with full configs are welcomed.
More from Multimodal
- Luma shares filmmaking-style workflow: motion-capture real stunts then restyle with video-to-video — lmoroney · 2026-10-18
- MOSS-Image-SFT at 4 Steps Draws Hype: 320x320 to 2K Without Artifacts — Powerful_Evening5495 · 2026-10-18
- MOSS SFT vs Flux 3 Klein vs SD 3.0: Same-Prompt Comparison Exposes SD3's Age — ZootAllures9111 · 2026-10-18
- Dev builds Mumbai-inspired 3D manga background generator from OpenStreetMap data — CurieuxExplorer · 2026-10-18
- Omni 1'1 flash video demo generates atmospheric foggy Victorian haunted house — Dense-Instruction210 · 2026-10-18
- Flux 3 klein distilled tested: int8 workflow delivers surprisingly good 2MP images — Ok_Tale7582 · 2026-10-18