Two cats and two dogs
The four benchmark subjects varied species, fur, pose, and setting so the launch choice was not based on one flattering pet image.
PetShimmy
AI and quality methodology
How PetShimmy turns a pet photo into a reviewed dance video.
The current launch system uses a pinned DreamActor M2.0 model through Replicate for motion transfer. PetShimmy controls the movement, validates the file, adds the exact matching audio, and keeps the result private until a human review approves it.
Method reviewed 31 July 2026
Motion transfer, not text-only video generation.
- DreamActor M2.0 via Replicate
- Exact-version and motion-asset controls
- Human approval before delivery
The production path
Six boundaries between upload and delivery.
Each boundary has a different job: reject unsuitable input, control the motion, verify the provider output, assemble the audio, validate the final file, and decide whether a customer should receive it.
- 01
Validate and normalize the pet photo
PetShimmy accepts one portrait JPEG, PNG, or WebP. It checks the real file structure and dimensions, then prepares a normalized pet image for the model.
- 02
Resolve one exact silent motion reference
The chosen dance and purchased duration must match a registered motion asset. Its location, type, size, dimensions, timing, and SHA-256 are checked before use.
- 03
Run the pinned motion-transfer model
Replicate receives the normalized pet image and the verified silent reference for the pinned DreamActor M2.0 version. PetShimmy does not rely on a free-form text prompt to invent the choreography.
- 04
Accept only a valid video-only intermediate
The returned media must come from a trusted provider host and pass byte-size, MP4, duration, portrait, codec, and native-720p checks. Any provider-returned audio is discarded.
- 05
Add and validate the matching soundtrack
PetShimmy combines the video with the exact audio companion prepared for that dance and duration, then verifies video, audio, duration, track continuity, and timeline alignment.
- 06
Hold the private file for human review
A technically valid result is still not customer-ready. It remains private until a named reviewer checks subject identity, obvious dance, framing, temporal consistency, and the delivery-quality policy.
The launch benchmark
Eighty provider runs, manually reviewed.
The retained launch benchmark used five dances, four pet subjects, two durations, and two independent seeds: 5 × 4 × 2 × 2 = 80 DreamActor M2.0 outputs.
Every provider call completed
All 80 recorded calls returned an output and every output was manually reviewed. Completion is reliability evidence—not a claim that every clip was fit to deliver.
Catalogue approval stayed conditional
The five dances progressed to launch with mandatory finished-output review. Animal motion transfer can still produce anatomy or consistency defects on a new photo.
Why publish the caveat? A provider success rate and a customer-usable delivery rate are not the same measure. PetShimmy does not turn “80 jobs completed” into “80 perfect videos.”
Known generative limits
The dance is controlled. Every frame is not.
The model interprets human movement for an animal body. Timing, poses, expression, paws, limbs, fur, anatomy, and background details can change. Small variation is expected; a substantial defect is a quality issue.
Audio is deterministic
The soundtrack does not come from the model. PetShimmy uses the prepared companion for the selected dance and duration and verifies it against the video timeline.
Visual quality needs judgment
Automated checks can prove file integrity, duration, and tracks. A person decides whether the pet remains broadly recognizable and the dance is clear enough to deliver.
Judge the output yourself
Five dances. One example pet. No testimonial theatre.
Watch the actual six-second AI pet results and hear the matching audio before choosing a dance.