The web app is out of private testing. What shipped with launch:

• New landing page built entirely from real engine renders — the before/after slider, the pipeline walkthrough, and every close-up are actual outputs, not mockups.
• Recalibrated pricing on the v2 engine's economics: Free now includes 20 credits/month, Starter is $12 for 250, Pro $39 for 1,200, Studio $99 for 4,000. The one-off Launch Pack (300 non-expiring credits) dropped from $29 to $19.
• Batch translation: drop up to 10 files, pick your languages once, and follow the whole render matrix on a live progress board.
• The review screen is now a first-class surface — detected elements boxed over your source image, per-language tabs, inline edits, then approve to render.
• Google sign-in, alongside the existing email magic links.

The engine now understands what kind of graphic it is looking at. A domain profile tunes decomposition and typesetting per document type:

• Menus: dish names can stay in the source language while descriptions translate, price columns are detected and never touched, and list spacing is preserved.
• Posters and banners: display headlines keep their letter-spacing and case style; call-to-action pills are re-set without redrawing the shape.
• App screenshots: UI chrome, status bars and numbers are kept verbatim; captions and feature copy translate.
• Everything else falls back to a generic profile with sensible wrapping rules for body copy.

We retired the generative image-editing pipeline and replaced it with a deterministic engine. Instead of asking a model to repaint your graphic, v2 takes it apart and puts it back together:

• Text detection and clustering finds every element — headings, items, body copy — with its exact box, color, and style.
• The original text is erased into a clean background plate. When erasing over photographs leaves smearing, the engine detects it and auto-repairs the region with AI infill; anything it is unsure about gets flagged for review.
• Translations are typeset with a real font engine: matched typefaces, true diacritics, right-to-left shaping, and auto-fit so text never overflows its container.
• Same input, same output — every run is reproducible, and pixels outside text regions are carried over unchanged.