GlennHasABeard documents MiniMax H3 prompt failures in ComfyUI tests

• Character drift shaped the staging: GlennHasABeard hid most of the cat after finding that a whole cat without a reference image drifted off model.
• A shadow-transformation prompt initially produced an actual cat in the room, according to his teapot test.
• The four clips run for 15 seconds each, generated locally in ComfyUI and finished with Topaz to 4K, as his costume post documents.

Two yellow-green eyes and a tail tip supply the entire character in the jack-o’-lantern scene. MiniMax’s official model card lists output up to 2K, putting the creator’s 4K finish in a separate upscaling stage.

The clever part is the production design: each scene conceals the anatomy that would otherwise need to remain consistent.

1. Candy at the table edge. Prompt: “pushes one wrapped candy slowly across a table until it drops off the edge”. The camera is locked at table height. GlennHasABeard explains in the candy post that the pumpkin hides everything except the paw because a whole cat without a reference image drifts off model.

A paw pushes a wrapped candy off the table.

1. Eyes inside a pumpkin. Prompt: “two yellow-green cat eyes open inside it and look out through the carved eye holes”. The camera is locked at step height.

1. A sliding ghost costume. Prompt: “A small white cloth with two eye holes slides across a wooden floor by itself”. The camera is locked at floor height. The sheet conceals the cat’s body entirely.

Prompt: “the teapot’s shadow slowly turns into the shadow of a sitting cat”. The camera stays locked, slightly to the right.

Early generations put a real cat in the room and cast its shadow onto the teapot. GlennHasABeard describes the intended transformation as starting with the teapot’s own shadow, then changing its shape on the wall.

He reported three generations total for this clip. The post gives the successful prompt, but does not supply the wording used for each earlier attempt.

Alongside the clips, GlennHasABeard published two broader questions and a separate numerical experiment:

• Prompt length: he asked whether short or long prompts get closer on the first try.
• Persistent failures: he asked what models still cannot do regardless of wording.
• Tracker scoring: his separate tracker update compared two rules over nine mornings. The main rule averaged a 0.42K miss against the backup’s 0.54K; he retained both after the backup came closer that morning.

The official ComfyUI workflows distinguish two H3 checkpoints:

• fl2va: text-to-video and first/last-frame image-to-video.
• ref2va: reference-to-video, accepting up to nine reference images, three reference videos and three standalone audio clips.

Comfy-Org publishes repackaged H3 model files for ComfyUI. GlennHasABeard names the applications, but his candy post supplies no GPU, seed or step count.