Design notes for building a desktop pet that stays out of your way

Havey Bachand — Thu, 05 Nov 2026 07:53:08 +0000

Desktop pets are a fun corner of desktop software: a small character lives on top of your windows, wanders around, and reacts to what you do. I work on DesktopCat, a free desktop cat for macOS and Windows, and here are a few design considerations that come up for any app of this kind.

1. The pet must never get in the way

A desktop pet lives on the screen all day, so it has to be unobtrusive. That means a transparent window, no stealing keyboard focus, and clicks passing through everywhere except on the character itself.

2. Keep it light

Idle animation should cost almost no CPU or memory. A lightweight cross-platform shell (we use Tauri for the app) helps compared with shipping a full browser runtime, though sprite animation and window management are where most of the tuning goes.

3. Reactions make it feel alive

Small touches matter more than complex AI: napping when you are idle, looking up when you type, letting you pet, feed and drag the cat. Random-but-believable behavior beats scripted loops.

4. Be honest about what is free

DesktopCat itself and the default cat, Mochi, are free with no account needed. Optional extras (an official cats pass and a human-made custom cat from your own photos) are sold separately. The app is closed source.

Try it

macOS 13+ (Apple Silicon) and Windows 10/11 (x64): https://desktopcat.com/download

Feedback on behavior ideas for a desktop companion is very welcome in the comments.

How the "train window toy car becomes a real car" AI video trend works (and how to shoot a photo that works)

Havey Bachand — Mon, 02 Nov 2026 02:12:04 +0000

If you've scrolled short-video feeds lately you've probably seen it: a tiny toy car sits on the windowsill of a moving train, the train dives into a dark tunnel, and when it comes out the same car is full-size, driving down an open road. It looks like a practical effect, but almost every version you see is image-to-video AI.

I built a small tool around this one effect, so here's what I learned about why it works and what makes a good source photo.

Why the tunnel matters

Image-to-video models are good at short, continuous motion and bad at big scale changes inside a single unbroken shot. The tunnel solves that. The frame goes dark for a beat, which gives the model a natural "cut" where the scale of the car can jump from miniature to life-size without the viewer seeing the morph. The train window frame also helps: it's a strong, stable composition that keeps the model from drifting.

So the shot is really three beats:

Toy-scale car on the windowsill, landscape moving outside.
Tunnel darkness (the transition).
The same car, full-size, on the road.

What makes a good input photo

After a lot of test renders, these mattered most:

One car, clearly visible. Rear three-quarter angle works best because it reads well both as a toy and as a real car.
Daylight, simple background. Busy parking lots confuse the identity of the car between beat 1 and beat 3.
No people, no readable license plates. Faces and plates get mangled and are a privacy issue anyway.
Not too cropped. The model needs the whole silhouette to keep the car consistent across the tunnel.

Output format choices

For this kind of clip I settled on ~15 seconds, 720p, 9:16 vertical, with generated ambient sound (train, engine, road). Vertical because that's where the trend lives, and sound because a silent clip of a train feels wrong immediately.

The tool

If you want to try the effect without writing prompts, I put it at TrainWindowAI: upload one car photo and you get the MP4. It's a one-time $8.99 per video, no subscription, and failed renders return the credits. It runs on Seedance 2.0 Mini.

Happy to answer questions about the prompt structure or the photo tips in the comments.

DEV Community: Havey Bachand