Local MiniMax H3 tests show fire appearing despite negative prompts

• A matchbox pulled a local MiniMax H3 test toward fire, according to the creator's matchbox post, although his posted clip shows a breathing creature without a visible flame.
• Explicit exclusions shaped another scene: the water-bowl post says the prompt ended with “no hands, no clay.”
• Sound supplied timing in a flipbook bird and a crumbling toast clip, alongside locked cameras and short, specified actions.
• Seven marble shots took under three seconds each on a hosted service, while a 15-second local take approached an hour on the creator's PC, according to his marble post.

A moon-making test needed a fingernail of light instead of the word “crescent.” A spiral test needed one moving tip instead of a request for a spiral. Those tiny prompt revisions are the interesting part of this run of 15-second experiments.

GlennHasABeard described a matchbox as a cue that made H3 want to add a flame, even though fire was absent from his prompt. He put “no matches” at the end and asked for the creature inside to rise and fall “as if breathing.” The posted clip shows the creature; the flame problem appears in his account of the experiment.

There is a distinction between that in-prompt exclusion and a dedicated negative-prompt channel. MiniMax's open-weight README calls the released checkpoints CFG-distilled; a ComfyUI node builder says conventional negative prompting does not work with H3 and favors distinctive positive descriptions when separating subjects.

For a second local render, the creator described walls that “rise and thin into a smooth round bowl, as if shaped by invisible hands.” He appended “no hands, no clay,” saying pottery prompts otherwise pulled both into the scene. The 15-second take used a fixed camera at wheel height.

His local clips ran through ComfyUI and were then upscaled with Topaz to 4K, as the chocolate-rooftop post also specifies. ComfyUI's H3 workflow guide provides native open-weight templates for text, image and reference-to-video generation with jointly generated stereo audio.

The fixed camera leaves the prompt to assign motion precisely. Across the experiments, the instructions break down into small physical jobs:

• A grater holds still while chocolate shavings fall onto miniature roofs, according to the rooftop experiment.
• A knife travels down and up, never sideways through a block whose slices reveal tiny deserts, in the sand-cutting clip.
• Two cut lemon halves turn at the same slow speed around their galaxy-like interiors in the lemon clip.
• Six plants lean toward an unseen passerby and straighten one after another in the plant clip.

“Spiral” produced rings on the first take, the creator reported. His revision specified “ONE continuous line with ONE moving tip” that gets farther from the center as it travels. Another prompt sent a leaf's break line from stem to tip over nine seconds rather than letting it crumble all at once.

The moon experiment had a similar geometry problem: twelve pieces came out as a pile, then three as two halves and a full. Describing a dark ball with a thin glowing sliver along its right edge finally gave the model a third shape, according to the moon post.

The bird's wings change position as notebook pages riffle. The creator called for one continuous riffle sound to hold the flip rate steady, while the toast test used one crunch per second for ten seconds to count impacts and crumbs.

He gave five stacked stones five clacks and reserved quiet for the soap bubble landing on top, as the stone-stack post explains. In the sky-colored tape measure, ten seconds of retraction end with a click and a dimming bench. MiniMax's prompt-writing guide likewise separates shot-by-shot action from the overall physical soundscape and background music.

A separate marble sequence used one reference photo across seven individually prompted five-second shots, the creator said. There was no music; each sound came from the marble. In a reply, he confirmed the same photo was used for every clip.

Those shots were made on Reactorworld, where he reported each five-second clip was ready in under three seconds. His own PC took close to an hour for one 15-second H3 take, a sharp production difference for the same creator's short-form experiments.