Skip to content
AI Primer
workflow

GlennHasABeard fixes a reversed paw print in local MiniMax H3 clips

GlennHasABeard uses constrained prompts and locked cameras for local MiniMax H3 clips in ComfyUI. He reports fixing a reversed paw print, while generation takes about an hour per 15-second clip before Topaz upscaling.

4 min read
GlennHasABeard fixes a reversed paw print in local MiniMax H3 clips
GlennHasABeard fixes a reversed paw print in local MiniMax H3 clips

TL;DR

A marshmallow ghost succeeded where an earlier sheet-to-ghost attempt had only billowed. A spider dunked like a teabag initially went headfirst into the cup, then flipped underwater to climb out.

One press, one paw print

The first pie take touched down twice and left the print facing the wrong way. The gloriously fussy correction gets one touchdown and an impression aligned with the approaching paw, using a locked camera slightly above the pie.

The posted prompt excerpt: “leaves one clean paw print in the filling.”

The filling's displacement impressed the creator too.

Ears, tails and silhouettes

Three cat shots deliberately limit the details H3 has to maintain:

  • Ears and eyes: Each box flap gets its own second and side in the box shot, with only ears and eyes emerging over the rim.
  • Tail: A right-facing bag opening gives the tail a clear exit in the bag clip, shot side-on at floor height.
  • Silhouette: Removing fur and facial detail lets the creator show a whole cat without a reference, as he explains in the moon clip.

The box-height shot reveals only ears and eyes.

Transformations in place

GlennHasABeard repeatedly specifies where a transformation originates:

  • Soap curl → bat: The bat forms from the shaving on top of the block, with no new creature flying into frame, according to the soap-cutting thread.
  • Marshmallow → ghost: The change stays on the skewer in the roasting test.
  • Orange → jack-o'-lantern: The tiny pumpkin occupies the fruit's original position in the peeling clip.
  • Leaf → moths: The prompt identifies the moths as the leaf itself, rather than insects arriving from elsewhere, in the leaf test.

Zippers, fog and body orientation

The motion notes constrain travel paths as well as transformations:

  1. Zipper direction: The zipper runs left to right across the pumpkin's front in the pumpkin reveal, giving the pull one route before the upper half opens like a lid.
  2. Opening direction: The left half moves left and the right half moves right in another fog clip, giving the vapor a single outlet along the fold.
  3. Material behavior: Describing fog as pouring “like milk” keeps it moving downward in the pouring test, although the pour still misses the bowl.
  4. Body orientation: The revised spider setup holds the animal head-up throughout, dipping only its back end into the water.
  5. Travel endpoint: The broom prompt specifies one long stroke across the porch and off its open end in the sweeping clip.

Counts still slip

The crow-and-pumpkin prompt asked for one tap and one answering knock. The render produced several taps and two knocks; GlennHasABeard wrote that H3 “still won't hold a count,” while accepting the exchange as a successful conversation.

About an hour per 15 seconds

GlennHasABeard reports roughly an hour of generation for a 15-second clip. He queues a week's worth of tests, lets them run, then picks the takes he likes best.

The timing reply gives no GPU model.

Weights, resolution and commercial rights

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 5 threads
TL;DR1 post
One press, one paw print1 post
Ears, tails and silhouettes3 posts
Transformations in place4 posts
Zippers, fog and body orientation4 posts
Share on X