Skip to content
AI Primer
workflow

Creators test local MiniMax H3 in ComfyUI for 15-second motion renders

Creators used MiniMax H3 in ComfyUI to direct ordered weather changes, object transformations, and material movement in local 15-second renders. The tests also used audio cues and locked cameras, and several results were upscaled to 4K.

4 min read
Creators test local MiniMax H3 in ComfyUI for 15-second motion renders
Creators test local MiniMax H3 in ComfyUI for 15-second motion renders

TL;DR

MiniMax open-sourced H3 in August, and ComfyUI's official workflow guide describes open-weights text, image, and reference-to-video workflows with native stereo audio, up to roughly 15 seconds at 24fps. The fog jar turns its sound track into a counting device: three lid squeaks produce three screw-top rotations in GlennHasABeard's fog clip.

Local render baseline

These are text-to-video runs. GlennHasABeard wrote in the local setup post that styled starting frames drifted from their look before a clip ended, so the showcased experiments avoid image-to-video.

The stated baseline was:

  • ComfyUI with open H3 weights
  • 864x480 resolution, 24fps, 20 steps
  • no Turbo LoRA
  • res_multistep sampler with the simple scheduler
  • Topaz upscaling after generation

A 15-second run took about 50 minutes on the creator's RTX 3060, per GlennHasABeard's render-time reply. GlennHasABeard's Topaz reply identifies the local upscaler as Topaz Video's Proteus.

The bottleneck is moving too. ComfyUI's September 22 VAE update cut a 1344x768, 129-frame encode/decode round trip from 24.3 to 12.7 seconds, a VAE-only measurement rather than a full-generation time.

Shot lists and sound cues

The prompts work like miniature production instructions: identify the action, state the order, then pin the camera. The weather prompt in GlennHasABeard's weather test specifies white cloud, grey rain, snow, then orange sunrise, preventing the four-panel frame from collapsing into repeated weather.

The audio layer gets similarly literal treatment:

Consistency lines

GlennHasABeard calls out a short negative constraint as a “consistency line,” used where the default material behavior would ruin the shot.

Material behavior

The tabletop clips borrow familiar mechanics instead of asking for pure abstraction.

  • The ice shaving travels through the hand plane's throat and curls over its top, addressing earlier takes that dropped it in front of the blade in GlennHasABeard's ice-plane clip.
  • Golden light is written as thick liquid from a sponge, so it runs, pools, and finds a level on the counter in GlennHasABeard's sponge test.
  • A cut through sand exposes a horizontal water band with tiny breaking waves; its eight-second visual cut is matched to an eight-second hiss in GlennHasABeard's sand-cut clip.

Fluid typography

CharaspowerAI used the same material logic at title-sequence scale. The prompt starts with two ocean walls, runs the camera between and around their collision, then pulls the suspended water into the word “TIDAL” before a left-to-right pressure pulse detonates the letters.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 5 threads
TL;DR4 posts
Local render baseline2 posts
Shot lists and sound cues3 posts
Consistency lines2 posts
Material behavior1 post
Share on X