Two-pass workflow fixes flat AI acting and voice in MiniMax H3 video generation
infearia · reddit · 2026-11-09
A Reddit user shares a two-pass workflow (Pika H3 Flex + H3 reference model) to fix flat-sounding AI voices and stiff acting:
- Pass 1 (low-res draft): Add a pre-rendered first frame via the Add Guide for MiniMax H3 node, prompt only for action and emotion (e.g. "seething with rage") without exact dialogue, and generate at 0.3MP. The gibberish dialogue doesn't matter — the character already sounds and moves like a real human.
- Pass 2 (final render): Same workflow at target resolution (0.98MP), feeding the pass-1 frames and audio into the H3 Reference to Video node, with the final dialogue written in <d> tags.
Key gotcha: reference-video resolution needs balancing. At 0.3MP the model followed the reference too closely and lip-sync broke — the author downscaled to 0.1MP, at the cost of softer facial textures. Raise resolution if output is muddy, lower it if H3 starts copying the reference; expect to try different seeds.
More from Multimodal
- DIY vocal synth models the vocal tract as acoustic tubes, yodels impressively — TinfoilTricorn · 2026-11-09
- Grim Reaper megastructure AI art stuns with apocalyptic scale — LudovicCreator · 2026-11-09
- User generates a full anime music video locally with MiniMax H3 in under 15 minutes — No_Range_8261 · 2026-11-09
- Google's AI Mode lets users turn selfies into Diwali greeting cards — ctjlewis · 2026-11-09
- AI image gen fail: prompt for interior design produced a McDonald's-themed strip club — cyrus_zei · 2026-11-09
- Marketer hunting for pt-BR voice models: currently using MiniMax H3 for full video generation — New-Dealer-5557 · 2026-11-09