A creator tests literal prompts for MiniMax H3 clips in ComfyUI

• A spiral that first came out as rings became a single widening line after GlennHasABeard specified one moving tip, according to his spiral test.
• His 15-second local H3 clips run through ComfyUI and get upscaled in Topaz; the butter-to-moths clip shows how literally he describes each change of shape.
• Sound sets the clock: his forest-stamp post assigns a thump one beat and the ensuing growth nine seconds of rustle.
• For a seven-shot marble reel, his reply confirms he reused one reference photo; the reel post reports much faster hosted generation than his local takes.

His moon tumbler needed a fingernail-sized sliver of light before it yielded a third shape. His potted plants tracked an absent cat with six successive rustles. ComfyUI’s H3 guide says the open-weight model generates sound and picture together, which makes those audio cues part of the shot design.

GlennHasABeard describes a repeatable setup across these tests: a locked camera, a 15-second local H3 render in ComfyUI, then Topaz upscaling to 4K. For the tape-measure shot, he gives the tape a face made of evening sky, orange near the case and violet at the far end.

The 4K finish is an upscale. ComfyUI’s local workflow documentation puts H3’s native canvas at a 768-pixel short edge and offers text-to-video, image-to-video and reference-to-video templates.

The first take drew rings. GlennHasABeard rewrote the motion as “ONE continuous line with ONE moving tip” whose distance from the center keeps increasing, then held the camera overhead for the entire trace.

“Crescent” was too loose: one batch produced twelve moons as a pile, another produced two halves and a full moon. The prompt that gave him a third shape described a dark ball lit by a curved sliver “shaped like a fingernail clipping,” covering a tenth of its face.

The transformations specify what moves and what remains behind:

• Butter shavings fold in half, open as pale wings and lift off the board as moths in the butter clip.
• A stroke flips velvet’s nap dark behind the fingers; a seam of gold opens along that band in the velvet clip.
• In a separate Seedance 2.5 experiment, a cloud patch tightens over eight stitches while the blue sky behind it disappears stitch by stitch. The mending post says this used one Prism reference, a 480p render and Topaz upscaling.

MiniMax’s H3 model card specifies native stereo audio. GlennHasABeard writes sound into the action rather than leaving it as atmosphere:

• A lemon produces glass beads: a wet squeeze and a glass tinkle tell the model the two actions overlap in his lemon post.
• A stamp lands with one thump, followed by nine seconds of forest-growth rustle in his paper-forest post.
• A pottery wheel’s hum rises and slows over eight seconds, while rain is confined to the bowl’s rim in his wheel post.
• Six plants follow an unseen presence, one rustle per plant, in his plant post.

For a reel made on Reactor, each five-second shot had its own prompt but used the same photo of a marble, as GlennHasABeard confirmed in a reply. He says each hosted clip was ready in under three seconds; on his PC, a single 15-second H3 take runs close to an hour. The shots and hardware differ, so those times describe his two production paths rather than a like-for-like speed test.

Reactor’s Sol-H3 announcement describes a speed-tuned FastH3 variant running on Blackwell hardware. GlennHasABeard’s reel post does not identify the variant used for his clips.

The production trail continues after rendering. GlennHasABeard says his ComfyUI outputs are labeled, so he can ask Claude to make a CSV for Zernio’s scheduler; he had Monday and Tuesday’s clips queued while continuing to post to X manually, according to his workflow reply and his scheduling update.

He also disclosed that his Zernio link is a referral link and that he earns a share of referred customers’ payments.