Pinned Tweet
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
Nate O retweeted
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
Click a label, say the fix. Meta's speech model hears it, SAM 3.1 outlines what you clicked, and 21 labels go from failing contrast to passing in one checked change. @MetaforDevs @AIatMeta teapawt.net/precogpunk/status/2117…
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
Hi all, I'm Nathan and I build design tooling. I compared a pixel-based contrast check with axe-core on 28 sites: it agrees on 99.9% of verdicts and settles 97% of the ones axe leaves for manual review. Then the fix is one token change. Would love this crowd's take on the gaps. teapawt.net/precogpunk/status/2117…
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
From a screenshot to a design system: every color in the app named, organized into three tiers, and ready for dark mode. Built for the Meta AI hackathon, open source. teapawt.net/precogpunk/status/2117…
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
Voice in, checked change out: speech to text, SAM 3.1 outlines what you clicked, the fix lands as a design token, then every text pair on the page is re-checked. Contrast fixes are computed, not guessed by the model. teapawt.net/precogpunk/status/2117…
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
Most apps I inherit have no design tokens, so nobody can change a color safely. This gives an app a reviewable token set from one screenshot. Then dark mode or a seasonal theme is a token change. teapawt.net/precogpunk/status/2117…
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
The model proposes, code decides. Muse Spark names the design tokens, deterministic checks validate every name, and four pixel checks have to pass before a PR opens. On 28 real sites, 0 pixels changed. Open source: github.com/nhunsaker/screens…
Three Meta models in one tool: SAM 3.1 finds the buttons from pixels, Muse Spark names every color, and Meta's speech model takes the fix by voice. Runs on open weights. Built for the Meta AI hackathon. teapawt.net/precogpunk/status/2117…
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
Rewrote 3,687 hardcoded colors to CSS variables across 28 real sites. Zero pixels changed, checked on every one. The originals stay as var(--token, #hex) fallbacks, so nothing breaks. Open source. teapawt.net/precogpunk/status/2117…
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
Building in public: retrofitting design tokens into apps that have none. A screenshot goes in, SAM 3.1 finds the controls, Muse Spark names every color on three tiers, and the rewrite must change 0 pixels. What would you check before trusting it? teapawt.net/precogpunk/status/2117…
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
3
7
138
Reading a page's code finds 0 of its 8 buttons. SAM 3.1 open weights on a @modal L4 finds all 8, and it matched hosted SAM on every test. It's the eyes of Screenshot to Tokens. teapawt.net/precogpunk/status/2117…
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
5
117
Muse Spark Open on @baseten names every color in an app on three tiers, the raw value, what it means, where it's used, in 3 to 5 seconds. That speed is what makes Screenshot to Tokens feel instant. Thanks for the hackathon credits. teapawt.net/precogpunk/status/2117…
Built for the @AIatMeta Global AI Developer Hackathon: Screenshot to Tokens.
A design system from a screenshot in seconds. A more accessible app from the tokens in one change.
Try it live, click a label and talk to it: s2t.sorbcloud.com
Open source. Thread below.
4
124
Replying to @precogpunk @AIatMeta
Every number in the video comes from a real run, and the runs are in the repo. Apache 2.0, clone it and run it on your app today.
Code: github.com/nhunsaker/screens…
Video: youtube.com/watch?v=RpnO9HfF…
Built with SAM 3.1, Muse Spark Open on @basetenco, Meta speech. @cerebral_valley
Built on Meta's models with partner credits from the hackathon: SAM 3.1 open weights on @modal_labs, Muse Spark Open on @basetenco with @FireworksAI_HQ as the failover, and Meta's speech model for the spoken requests. Thanks @MetaforDevs @AIatMeta @cerebral_valley.
8
1,024


