Why AI screenshot-to-code tops out at 49% fidelity

Ask an AI to turn a screenshot into code and you get something that *looks* plausible and is almost never pixel-accurate. The reason is architectural, not a quality bug you can prompt away.

A vision model sees flattened pixels — not the DOM, not the exact widths, hex colors, font names, or the real image and SVG files. Every measurement is inferred. Worse, vision transformers tokenize the image into patches, which destroys fine-grained spatial information: the model knows a button is 'bottom right', not that it's at x=1192.

The leading academic benchmark (Design2Code) found top multimodal models matched the original page in roughly half of cases, while frequently omitting real elements and inventing wrong ones. 'Looks like a reasonable webpage' is not the same as 'reproduces this webpage'.

FrameFlow takes the opposite path: a real browser resolves the site into exact pixels, we capture that ground truth, and we prove the match with a per-region pixel diff. Deterministic capture first; AI is never in the fidelity path.