MiniMax H3 fails to precisely replicate video with detailed prompt
unraveleverything · reddit · 2026-08-19
A user tested MiniMax H3's controllability by attempting to force a precise, no-change replication of a source video using highly structured prompts. Despite explicit instructions to preserve subjects, video, and audio exactly, the model hallucinated content and became confused temporally. Various configurations (attention types, samplers like Euler, steps) were tested without success.
More from Multimodal
- Seedance 2.5 video explores cinematic color and composition — Eric520CC · 2026-08-19
- Krea AI Hints at Incoming Krea3, New Image Generation Version — listopalafoto · 2026-08-19
- AI film 'Teo' streams 24/7 on Roku and LG via Fairground AI Creator TV — DavidmComfort · 2026-08-19
- Office sitcom "Still Loading" Ep 5 fully generated with MiniMax H3 — asankhs · 2026-08-19
- Human + Claude + Blender Built 3D Video in 4 Weeks: $20k Thinking Cost — majidmanzarpour · 2026-08-19
- Minimax video test: Flux 3 level visuals without audio — mesmerlord · 2026-08-19