Reve 2.0 releases with 4K image output and layout control

• Reve 2.0 shipped with 4K output and a new layout-first control system that, according to the launch repost, can generate and edit images through a precise intermediate representation instead of prompt-only steering.
• In The Layout Bet, Reve says the new model is built around a "Large Layout Model" that turns instructions, layouts, and images into a structured scene description before rendering pixels, a claim echoed by the architecture repost.
• The creator-facing pitch is tighter control over composition, with the launch video showing clean layout placement and a follow-up repost framing the system as image generation and editing "from code."
• Early leaderboard traction arrived fast: the Arena repost says Reve 2.0 landed at No. 2 in Text-to-Image Arena, and the live leaderboard lists Reve at 1280±11 from 3,455 votes.

You can try the model on Reve's site, read the company's full architecture note in The Layout Bet, and compare its early standing on the live Text-to-Image Arena leaderboard. The interesting bit is not just sharper images. Reve is betting that image generation should act more like structured design software, where layouts, regions, and local edits are first-class controls.

Reve's big claim is that text is too fuzzy for reliable composition. In The Layout Bet, the company says most image models rely on long-form text as the internal plan, while Reve 2.0 swaps that for a layout that stores location, size, local description, color, and optional image references for each element.

That layout acts like a scene backbone. Reve compares it to HTML or SVG for images, and says users can refine results either with natural-language instructions or by directly editing the layout structure.

The architecture change is more than a UI wrapper. According to The Layout Bet, Reve built a unified "Large Layout Model" that accepts any mix of layouts, instructions, and images, derives a layout during its internal reasoning, then renders the final pixels.

The company also says it trained the system on billions of images with dense human annotations, then continued pretraining and post-training open-source language models for spatial reasoning around its layout representation. The architecture repost describes Reve 2.0 as a completely new architecture built from the ground up.

The workflow Reve is pushing looks closer to design iteration than one-shot prompting. The research post says reconstruction quality improves as the model gets more image regions to work with, and that layouts become more useful for targeted edits when source pixels are provided.

That matches the launch framing in the launch repost and the control repost, both of which emphasize generation and editing through a precise, code-like representation. On the consumer side, Reve's homepage showcases branded label placement, object swaps, and composition tweaks as core tasks rather than side features.

Reve 2.0's first public score gives it immediate company in the frontier pack. The Arena repost says the model landed at No. 2 in Text-to-Image Arena, above Nano Banana 2 and MAI-Image-2.5, and the live leaderboard currently shows Reve at 1280±11 with 3,455 votes, just behind OpenAI's top entry and narrowly ahead of Google's neighboring model.

That does not settle the longer-term ranking race, but it does give creators a concrete comparison point on day one: Reve is entering as a layout-control play that is already scoring near the top of the public image-model stack.