newsbubbles/flowcard

Flowcard helps AI language models understand motion in video, even though those models cannot actually watch video. It takes a video clip and compresses it into a single annotated image, called a flowcard, along with exact numeric stats, so a model that is strong at reading images but blind to video, like Claude, can spot things such as game glitches, stalled rendering, sudden teleports, flickering surfaces, and timing problems in animation. The idea borrows from a technique from 1878: collapsing motion into still images that can be studied. The reasoning behind the project is that even video capable AI models are still weak at judging direction of movement, so flowcard sidesteps the issue entirely by using traditional computer vision to render motion spatially in a single image, which the AI can then read using its stronger image skills. No AI model runs inside the tool itself, so it is free, local, and produces the same result every time. Each flowcard is a single image with six labeled panels, plus a timeline and a strip of frames, covering things like which pixels changed, when motion last happened in each area, the direction and speed of movement with camera motion removed, flicker patterns, and tracked object paths. A separate stats file carries the same information as exact numbers. Flowcard works by first estimating camera movement between frame pairs, then stabilizing the video into a consistent coordinate space, and finally running several detectors on top of that stabilized footage. It ships as tools that connect to AI agents through MCP, plus command line equivalents, and includes a validation script that checks its output against clips with known, planted problems. A benchmark included in the project shows that giving a language model the flowcard and its stats, instead of raw sampled frames, leads to far more accurate, measurement based answers about what happened in a video.