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.

Original post →

More from Multimodal

Multimodal channel →