There’s an interesting release at Hugging Face: screenshot2html. You hand it a picture of a web page and it hands you back the HTML. We’re finally reaching the point where a Figma mockup of a website can be converted, quickly, into working HTML by someone with no coding experience at all. That could help people launch ideas faster.
Some context on what’s under the hood. The Space runs a model its makers call VLM_WebSight_finetuned, built on Mistral-7B and SigLIP and trained to convert “screenshots of website components into HTML/CSS codes.” It learned from WebSight, a set of 823,000 web pages paired with their screenshots — all of them synthetic, generated by a language model rather than scraped from the live web. Synthetic data means clean pairs: every screenshot matches its code exactly, which is what you want when teaching a model to go from one to the other. The model card is candid that this is “very much an alpha version.” It is an alpha that already does the thing.
A mockup was always something a founder could make alone. The code was the part that needed a hire, or a co-founder, or six months of evenings. If the picture is the spec, then drawing it well becomes the job.
Next step — Figma, or PowerPoint, or whatever you happen to sketch in — to a fully coded web or mobile app, front end and back end both. Within 12 months? The front end is the easier half. A back end has to hold users and their money, and keep holding them next week, and that is a harder thing to infer from a screenshot. But the direction is set.
Anyone can launch any new digital business they can conceive of. And so can AIs directed to do the same. How many launches per minute?
