Reve launches 2.0 image model
Reve has released Reve 2.0, a native-4K image model that builds every image from an editable layout and debuts at #2 on Arena's text-to-image leaderboard, according to the company.
Reve announced Reve 2.0, calling it its most capable model so far. The company says the model debuts at #2 on Arena's text-to-image leaderboard, 125 Elo points above its v1.5 predecessor. Reve also says it trained the model on 10x fewer GPUs than competitors, though it did not name those competitors or give absolute compute figures.
The main technical change is what Reve calls "images you can touch." Each image is generated from a layout, which the company describes as a structured map of the objects, text and regions in a scene. Users can move a subject, rewrite the text on a sign or swap a background, and according to Reve the rest of the scene adjusts around the edit.
Other claimed improvements:
- Native 4K output that Reve says holds up at full size and is print-ready.
- Text rendering for signs, labels, posters and logos, which the company says comes out clean where other models struggle.
- More reference images per composition, so users can lock in a style, character or look.
- A rebuilt editor that shows controls when an object is selected and streams results while the model works.
Reve 2.0 replaces all earlier Reve models and is now the default at reve.com. The company describes layouts as a first step toward treating images "more like code," meaning something humans and AI can read, write and reason over together. It points readers to a separate technical write-up, "The Layout Bet."
The announcement does not cover pricing, API availability, model size or how the layout system is implemented.