Reve introduces layout-based image model Reve 2.0

Anyone who has spent time with AI image generators knows the frustration: the scene is nearly perfect, but when you ask to move a coffee cup to the left, the model redesigns the entire room and changes the character's face. The culprit isn't necessarily the renderer—it is text itself. Language is inherently ambiguous, and ambiguity destroys precision.

Reve is tackling this problem from a completely different angle. Instead of using text as the internal representation between intent and pixels, the team unveiled Reve 2.0, an image generation model built entirely around structured layouts.

Ditching text prompts for blueprints

In standard image pipelines, an LLM expands a short prompt into a lengthy text description, which a diffusion model turns into pixels. Reve threw out that English prose intermediate step. In Reve 2.0, an image is represented as a structured layout where every element has an exact position, size, local description, and optional attributes like color or reference images.

The team compares this layout to HTML for a webpage or SVG for vector graphics. It separates semantic intent from pixel rendering, giving both humans and AI agents a readable, code-like canvas to manipulate.

To make this work, Reve trained a unified Large Layout Model capable of visual understanding and generation. Leveraging open-source Qwen models for continued pre-training and post-training, the team taught the LLM spatial reasoning over layouts. The training pipeline was trained on billions of images bootstrapped from dense human annotations.

Doing more with 10x fewer GPUs

Reve claims Reve 2.0 is the top-performing image generation model built by a sub-$1T company. Even more striking: they claim it was trained using 10x fewer GPUs than competitors.

According to Reve's ablation studies, layout-based models consistently beat prompt-based generators of identical size. In image reconstruction tests, text prompts inevitably drift away from the original image regardless of how detailed they are. Reve's model, however, was able to reconstruct fine details without a single input pixel simply by scaling up the number of regions.

The team also observed that scaling laws hold true here: increasing the number of output regions effectively expands the model's visual thinking context, yielding sharp improvements in output quality and enabling pinpoint image editing.

Why it matters

For designers, illustrators, and developers, Reve 2.0 addresses the slot-machine nature of generative art. In professional workflows, client edits are mandatory, but traditional diffusion models turn minor tweaks into complete redraws. A structured layout bridges natural language and spatial coordinates, letting users lock down specific regions while swapping out others.

For the AI landscape, Reve shows that clever architectural bets can still outmaneuver raw compute. While trillion-dollar giants throw massive clusters at brute-force prompt understanding, treating visual generation as program synthesis proved to be dramatically more compute-efficient. Relying on open-source Qwen weights also highlights how accessible frontier spatial reasoning has become.

That said, we think the real challenge will be user experience. Layouts are great for machines and power users, but ordinary creators do not want to write JSON or fiddle with bounding boxes all day. Reve will need an exceptionally fluid interface if it wants this blueprint approach to replace the frictionless simplicity of plain-text prompting.