Skip to content
AI Primer
workflow

GlennHasABeard uses local MiniMax H3 to animate hidden cats, then upscales to 4K

GlennHasABeard shares prompts that constrain a hidden cat’s movement in locked-camera clips using local MiniMax H3 in ComfyUI. Topaz upscales the footage to 4K, and one render reportedly takes 50–60 minutes.

3 min read
GlennHasABeard uses local MiniMax H3 to animate hidden cats, then upscales to 4K
GlennHasABeard uses local MiniMax H3 to animate hidden cats, then upscales to 4K

TL;DR

  • Hidden-cat animation needs surprisingly little cat: GlennHasABeard's hat clip coordinates a hat and tail without naming the animal in the prompt.
  • The 15-second clips are generated locally with MiniMax H3 in ComfyUI, then upscaled to 4K with Topaz, as GlennHasABeard describes.
  • Reported turnaround is 50 to 60 minutes per clip, according to GlennHasABeard, who shared a desktop with a 12 GB RTX 3060.

A bottle wandered around a cookie, so GlennHasABeard removed it from the icing shot. After a glow stick lit, went dark and relit, he rewrote its prompt so the light could only get brighter, then hold steady.

Hidden cats, three routes

The hat clip's best trick is directing a cat the prompt never names. Across three clips, GlennHasABeard gives the visible parts tightly bounded movements:

  • Hat: “lifts a little and slides slowly away to the right, as if something underneath is walking”
  • Paw: In the ghost clip, “a black paw with white toes pokes out through one of its eye holes”
  • Tail: “curls around the pumpkin's stem, holds, and slides away back down behind it”

The paw gets one opening to emerge from and retreat into; the tail returns behind the pumpkin. GlennHasABeard also says the locked camera matters for his ASMR-style clips.

Cutting out failed actions

Several finished shots came from removing something H3 struggled to depict. GlennHasABeard describes three substitutions:

  • Apple slicer → already-cut apple. H3 didn't know what the slicer looked like, so he removed it. The apple only has to fall open around a lit candle.
  • Bursting water balloon → dissolving bath bomb. He says the model couldn't fake the balloon's physics; the replacement prompt asks a pumpkin-shaped bath bomb to fizz away while turning the water orange.
  • Melting ice → a crack releasing chocolate. After two unsuccessful ice-melting attempts, he switched to one crack opening and molten chocolate slowly pooling below it.

Counts still drift

For the paper ghosts, the prompt asks for a chain pulling open from left to right. GlennHasABeard deliberately leaves out the number of ghosts and defines the movement across the frame instead.

An explicit count fared worse in the cat-and-pumpkin clip. He reports asking for one pumpkin and one push, then getting two of each.

RTX 3060 and overnight renders

GlennHasABeard describes his setup as mostly default ComfyUI and says he typically lets the longer runs go overnight. His hardware screenshot supplies the machine details behind the reported turnaround:

ComfyUI publishes native H3 workflow templates. MiniMax's local deployment guide identifies their model files as pruned and quantized Comfy-Org conversions and leaves minimum GPU memory unpublished.

The official ComfyUI H3 guide specifies native output at a 768-pixel short edge and 24 fps. It also says commercial use of locally generated outputs requires a MiniMax commercial license, available through Comfy.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 4 threads
TL;DR3 posts
Hidden cats, three routes2 posts
Cutting out failed actions2 posts
RTX 3060 and overnight renders3 posts
Share on X