Turning GPT GIFs into a cat film: what smoothing fixed, and what it didn’t

A cat crosses a room, jumps onto a windowsill, watches a bird, curls up and falls asleep. That became a 57.2-second silent edit assembled from short GPT-generated GIFs.

The useful experiment was smaller: take three selected shots and compare their native timing with a locally finished version. Keeping the cat’s size and position consistent across cuts turned out to matter as much as making its movement smoother.

The comparison

Watch the full 7.68-second comparison.

Both panels use the same three generated WALK, JUMP and CURL takes, at the same playback speed.

Sixteen drawings at 160 ms give 2.56 seconds per shot. Three shots give 7.68 seconds. The native drawing cadence is 6.25 drawings per second; the MP4 comparison is encoded at 25 fps. An export frame rate alone does not add motion. Interpolated frames estimate intermediate states between source drawings.

Watch the complete 57.2-second cat edit. That longer film includes other beats and editing decisions; the three-shot comparison above is a separate, shorter artifact.

What improved

The finished side combines several changes. Framing cleanup reduces visible scale and room changes, compositing stabilizes the background, and selective interpolation fills some of the gaps between poses.

That means this comparison cannot isolate how much improvement comes from interpolation alone. Reframing and compositing visibly contribute to the result. A cleaner way to measure that contribution would be a separate comparison with framing held fixed; that test is still to do.

What still needed a different take

During development, the cat disappeared around a cut, the bird’s contact with its branch was wrong, and the bird faded away instead of flying out of frame. Other revisions introduced scale changes or extra stripes on the cat’s back.

Those failures made frame inspection necessary. Additional intermediate frames cannot establish an action that the generated source never depicted correctly. Around a fast jump, interpolation also needs review for doubled contours and ghosting.

For this workflow, each replacement shot needs two checks: its beginning against the previous shot’s ending, and its ending against the next shot’s beginning. A take can improve the action while making either cut worse.

The resulting review questions are concrete:

• Does the cat keep the same apparent size, markings and silhouette?
• Do its feet, the floor and the windowsill agree about contact?
• Does the action progress through the jump, landing and curl?
• Does the sleeping shot contain visible breathing rather than a frozen pose?
• Do both neighboring cuts preserve position and direction well enough for the intended edit?

These are human review questions, not automated acceptance scores.

What is reproducible from the public code

The accompanying GIF Cinema repository and agent skill document shot planning, short GIF generation with still-image references, frame inspection, explicit retakes and MP4 assembly. I maintain that project.

The included three-shot astronaut example can be assembled without an API key. It is a separate native-timing baseline, not a reproduction of the finished cat. New generation requires the user’s own API billing and model access.

The extra compositing, interpolation and Studio interface used for the cat are private and are outside the public skill. The repository uses a source-available evaluation license, with commercial use, redistribution and hosting requiring permission.

Matching decoded frame pairs and the comparison notes make it possible to inspect the selected footage without publishing private prompts or request logs.

What I would test next

The selected edit still contains stepped source poses and small matte or shading differences. One cat film does not establish reliable continuity across other characters, styles or longer films.

The next useful test would hold framing and source takes fixed, compare native timing against interpolation, and review the jump at full speed as well as frame by frame. That would separate motion improvement from composition repair.

The lesson from this experiment is that a coherent little film needs both: acceptable source action and cuts that preserve enough visual continuity for the viewer to follow it.

Disclosure: AI drafted and edited this article from my experiment notes and the referenced artifacts. The animations are AI-generated; the comparison includes local postprocessing.