MiniMax H3 & Wan 2.2 — Local Video Generation

Two local video workflows ran on ComfyUI for Windows. Both clips are shown with their exact frames, resolution, timing, and audio state — MiniMax H3 followed the walking action with audio; Wan 2.2 held a coherent neon scene with a wheeled subject.

Four sampled frames of a red biped robot walking through a rainy neon alley

model / workflow

MiniMax H3 local pipeline + Wan 2.2 TI2V 5B FP16

task

advanced-local-video-observation-v1

checks held

observed

measured attempts

best rate / timing

30.4 s / 60.4 s

attempt ledger·every attempt shown

attempt	total	load	rate	checks	what happened
MiniMax H3	30.4 s	—	56 frames	Observed	608×352 at 24 FPS; walking action and native audio present.
Wan 2.2 5B	60.4 s	—	81 frames	Observed	832×480 at 24 FPS; coherent scene, silent wheeled subject.

MiniMax H3 followed the action

The 56-frame MiniMax H3 clip kept a recognizable red biped moving through a rainy neon alley. The sampled frames show the stride advancing while the subject and camera relationship remain readable.

The MP4 also contained native audio. Rain, servo, and city sound made this the stronger demonstration when sound was part of the requested artifact.

Wan 2.2 held the scene with a wheeled subject

The 81-frame Wan 2.2 clip preserved the neon alley, reflections, and a stable subject across the sampled frames. Temporal coherence was present.

The requested walking robot rendered as a wheeled robot, and the clip was silent. That is a difference in prompt adherence, not a runtime problem.

Why these aren't compared head-to-head

The workflows used different resolutions, frame counts, sampling work, output features, and model families. Comparing their server times as if they were the same race would be misleading, so both are shown side by side instead.

Inspect multiple frames before playing the clip.
Score motion coherence separately from prompt fidelity.
Record audio presence as an output feature.
Publish the exact workflow settings beside every time measurement.

Four sampled frames of a wheeled robot in a coherent neon alley — Wan 2.2 held the scene across 81 frames, but the requested walking action became a wheeled subject.

MiniMax H3 · 56 frames · 608×352 · 24 FPS · native audio present.

Wan 2.2 TI2V 5B · 81 frames · 832×480 · 24 FPS · silent output.

reproduce / inspect·downloadable evidence

Measured results →

Dimensions, frames, FPS, server time, and sampled telemetry for both runs.

MiniMax H3 MP4 →

The exact 56-frame, 608×352 clip with native audio used in this observation.

Wan 2.2 MP4 →

The exact 81-frame, 832×480 silent clip used in this observation.

Scope boundary

These are unlike workflows and are intentionally not ranked against each other. The report describes two observed runs and makes no hardware purchasing recommendation.