MiniMax H3 multi-subject video bug: Voices bleed between different reference characters
Wide-Researcher583 · reddit · 2026-08-10
Users have discovered a significant flaw when using the Ref2VA model of MiniMax H3: voices 'bleed' between different subjects when processing separate reference images containing multiple characters.
Specifically, while the visual identities of different subjects are perfectly preserved, the audio branch blends the voices or one subject's voice identity leaks into another. Users have tried various methods such as changing seeds, shortening clips, and explicitly defining subjects, but no reliable solution has been found yet.
Related event: MiniMax H3 Video Model Tests Reveal Multiple Audio and Motion Bugs(4 posts)→
More from Models
- willcb: Opus 5.5 doesn't smell bigger, it smells like they fixed the data — willcb · 2026-10-16
- Unverified claim: Apple and Meta's 816B-parameter Muse Spark Open runs on two Mac Studios — TansuYegen · 2026-10-16
- Gemini 4 model card: model eliminated months of trial-and-error on bioweapons research — Puzzleheaded-King584 · 2026-10-16
- Every AI picks 7: Anthropic engineer's rule for escaping model taste convergence — rubenhassid · 2026-10-16
- Brain teaser: strip all attention, MLP and embedding matrices from a Llama 3 — how many trainable params remain? — Sauers_ · 2026-10-16
- Lean v4.35.0 ships verified kernels — one written and proven consistent by Claude — geoffreyirving · 2026-10-16