Generating an Animated Short with Code: Claude's Taste Is On Point This Time

Someone used Claude to generate a Memphis-style animation. Every object in the frame has a deep indigo outline and hard shadow; saturated primary colors and pastels are laid down as large blocks, overlaid with dots, stripes, and zigzags. The complete method they shared is: write a deterministic frame(i) function with Canvas; it takes a frame number as input and outputs an image; scenes are generated from coordinates and a fixed random seed; export frame by frame with headless Chrome (a browser without a UI), then composite with ffmpeg (a video compositing tool); the soundtrack is also synthesized in code—a light pentatonic melody plus ambient sound.

This style, when applied to a specific scene, was set by someone to Beijing's East Third Ring Middle Road, passing the CCTV Headquarters. The camera slowly moves forward along the road, letting the CCTV building's signature tilted twin towers and cantilevered structure gradually enter the frame, with CBD high-rises, overpasses, traffic, bus stops, and street trees spread around it. On a clear early-autumn afternoon, vehicles pass in an orderly way, and cloud shadows slowly sweep across the building. A bright yellow bus drives past the camera, just revealing the full CCTV building behind it; then a falling leaf is swept up by traffic, briefly drifts along with the camera, and falls toward the roadside. The image is brisk and bright, capturing the everyday feel of a busy Beijing street.

The video below is a demo in the same style.

Eian posted two pieces; essentially they turn maps or street scenes into moving miniature models: first draw the scene, then let the camera move slowly, accompanied by small movements within the frame. One is spring fields on the Chengdu Plain, the other is autumn on Yulin Road in Chengdu; they said they were once again stunned by Claude's aesthetic sense. The rhythm of both animations is carried by small movements—birds flying past, cloud shadows sweeping by, fish swimming, people walking, cars driving. At the bottom of the frame there are also three centered lines of small text: title, season description, and latitude/longitude, changing with the camera.

The method isn't complicated. The core is the frame(i) function: input a frame number, output an image. Random numbers are reseeded by frame number to ensure different render batches match up. After everything is drawn, use ffmpeg to composite at the frame rate. The motion in the animation comes from a one-take slow push and rotation; the "story moments" in the frame—birds flying, fish swimming, leaves being swept up—bring the static map to life.

But there are several places in this workflow where things can go wrong. A netizen reminded: if you use Math.random inside frame(i), you must reseed by frame number, otherwise different render batches won't match. Font textures must finish loading before the first frame. The latitude/longitude line changes every frame; fixing the character slots to a width prevents the text from jittering left and right. Another netizen added: calculate time within a frame as i/fps, then specify -framerate when compositing with ffmpeg, so later if you want to change duration or frame rate you don't have to rewrite the animation.

One reply put it quite accurately: if a human made this, we'd be amazed at how talented that person is; now that AI makes it, we'll gradually treat it as standard practice. Tools have lowered the barrier, but taste is still scarce. The most moving parts of these animations are actually the small movements in the frame—a falling leaf swept up by traffic, cloud shadows slowly sweeping across a building. That's judgment, not parameters.