Minimax H3 test round 2: dual-POV dialogue near-perfect, first-person sports physics fails
lazyspock · reddit · 2026-08-24
Following strong interest in part one, the author ran a second batch of Minimax H3 video tests with full prompts included:
- Dual-POV dialogue (near perfect): a side-by-side of a man and woman facing each other, each shown frontally as the other would see them, with scripted dialogue in <d> tags, a hand-on-cheek gesture, and a 00:00–00:10 timeline. Impressive result; framing kept drifting to 3/4 despite prompts for frontal view.
- First-person ping-pong (very good): a full POV rally from serve to winning shot and crowd cheering; couldn't eliminate the fisheye effect.
- POV bathroom selfie (near perfect): shot from behind with the phone screen and mirror reflection both rendered correctly.
- POV soccer free kick (bad): best of three attempts, physics still don't work and fisheye returns.
Takeaway: structured prompts (timeline + <d> dialogue tags + soundscape fields) control camera POV and pacing fairly reliably, but first-person sports physics and fisheye distortion remain clear weaknesses.
More from Multimodal
- Halcyon early-fusion image models match late-fusion with 275x less pretraining compute, open-sourced — davidwromero · 2026-10-16
- Running Wan2.1 14B FP16 on a 12GB laptop GPU: VRAM swap stress test finishes in 4m11s — RYuu-1120 · 2026-10-16
- Striking Midjourney generations: black holes and planetary aesthetics — ciguleva · 2026-10-16
- Fable 5.5 may drop today as Andrew Curran hints at early release — kimmonismus · 2026-10-16
- Kling 4.0 Adds Timeline Editing, Endorsed by 'Seoul Spring' Director Kim Seong-su — xiaohu · 2026-10-16
- One prompt, no reference images: Argon builds 3D Gaussian splats from geometry equations in NumPy — fofrAI · 2026-10-16