consciousengines/h3-video-edit-showcase

MiniMax-H3 video edit showcase

30 edits across 5 source clips, generated with MiniMax-H3 ref2va. Each output is a 5.16 s / 124-frame segment at 832×480, produced from the source clip plus a hand-written six-section prompt. Nothing here is cherry-picked: every edit attempted is included, including the ones that did not work.

Generated on a single NVIDIA B200 at 3.5 min per edit — 105 min for the set. The same work takes 13–19 min per edit on an RTX PRO 6000 (sm_120), the difference being memory bandwidth (8 TB/s vs ~1.8).

What the model does well, and where it stops

category	clip	result
environment under a moving camera	road/	strongest. night, snow, autumn, storm and golden all transform the whole scene while road markings, the stone wall and the forward motion hold
camera flight	drone/	6/6, including altitude change and pitch
interior relighting	lounge/	6/6, cleanly differentiated; night reaches luma 44 from a 112 source
large body movement	person/approach, person/lean_in, crowd/stare_sprint	lands clearly
subtle gaze / expression / micro-behaviour	person/stare, person/stillness, crowd/freeze, crowd/look_up	weak — the honest limit

The pattern: spatial change of large magnitude lands; small changes do not. Someone walking into the lens works. The same person's expression shifting does not. Whole-scene relighting works. A crowd subtly changing its collective behaviour does not.

Two results that contradicted expectations set by earlier experiments on LongLive with the same footage:

• camera pitch works. LongLive tilt/high-angle edits "barely moved the framing"; H3 handled drone/dive and drone/climb_reveal cleanly.
• darkness works. "Evening never gets dark" was a persistent LongLive failure; lounge/night goes 112 → 44 luma and road/night 108 → 36.

Layout

lounge/   6 relight edits    + lounge_sheet.png
person/   6 behaviour edits  + person_sheet.png
crowd/    6 collective-motion edits
road/     6 environment edits
drone/    6 camera-flight edits
index.json   per-edit generation times

Each folder also holds _reference.mp4, the trimmed source segment actually fed to the model.

How the prompts are written

Six-section full-reference format per the model's own VIDEO_PROMPT_WRITING_GUIDE_ref_en.md, hand-written rather than generated. Three rules, all of them arrived at by measurement:

1. Additive phrasing only. "warm amber lamps glow" produces a dark room; "the lights dim" is ignored. Never describe what is removed.
2. Never name what the source does not contain. The model builds whatever the prompt names — an invented cabinet becomes a real cabinet.
3. Relight and motion need opposite scaffolding. A relight prompt asserts that every frame matches the source in geometry and camera. A motion prompt must not: that same language argues the requested movement away.

Generation code: showcase.py.