Reve 2.0 swaps text prompts for layouts
Image-generation startup Reve says Reve 2.0 makes images from a structured layout instead of an expanded text prompt, aiming to give users precise control over where things go and what color they are.
Reve detailed the approach behind Reve 2.0 in a post on its research blog. The company argues that today's image models depend on ambiguous text. An LLM expands the user's prompt into a long description, and a diffusion model draws from that, so a small change to the prompt can change the whole image.
Reve 2.0 uses a layout as the in-between step instead. A layout is a hierarchical description of the image. Each element has a location, a size and a local description, and can carry extras such as reference images or colors. Reve compares it to HTML for a webpage or SVG for a vector graphic. Users can change the result by writing instructions or by editing the layout directly.
The company says it built a unified "Large Layout Model." The model takes any mix of layouts, instructions and images, works out a layout as part of its reasoning, and then produces the pixels. Training data came from a pipeline covering billions of images, starting from dense human annotations. Reve then continued pretraining and post-training open-source Qwen LLMs to teach them spatial reasoning over layouts.
Reve reports an ablation in which layout models beat prompt-based generators of the same size. It also says quality improves both with model size and with the number of layout regions. It claims layouts can rebuild images more faithfully than text, even with no pixels as input. The post gives no benchmark numbers or outside evaluations. Its headline claim, that Reve 2.0 is "the best image generation model made by a sub-$1T company, trained on 10x fewer GPUs," has not been independently checked.
Reve calls layouts a first step toward treating image generation as program synthesis.