Two image generation systems, Ideogram 4.0 and Reve 2.0, are pointing to a new direction for AI image tools: structured, editable layouts rather than free-form text prompts alone. The shift reflects a broader effort to make generated images more controllable, especially when users need specific object placement, sizing, or composition.
Reve described its approach in a June 3 blog post, saying it replaced English prose as the main intermediate format with a layout. In this setup, each element in an image is represented with structured information such as position, size, a short description, and optional attributes like color or reference images. The company said this gives humans and AI agents a shared format that can be read and edited more precisely than a prompt.
According to Reve, text prompts alone are often too ambiguous to reliably direct visual output. Small wording changes can produce large changes in the final image, and specific instructions about where something should appear or what color it should be are hard to enforce through prose alone.
A layout-based system tries to address that by separating the image’s composition from its pixel-level rendering. Reve compared the concept to HTML for webpages or SVG for vector graphics, with the layout acting as a kind of backbone for the image. Users can refine results by editing the layout directly or by giving natural-language instructions that the model translates into structure.
The company said it built a unified model for both understanding and generating visuals. That model can take layouts, instructions, and images as inputs, then infer a layout internally before producing the final image.
Reve said it developed a data pipeline using billions of images and dense human annotations, followed by additional training on open-source large language models to improve spatial reasoning. The company did not provide a full technical breakdown in the public post, but said the method helped the system learn how to work with its layout representation.
The startup also argued that the approach can outperform prompt-based generators of similar size. It said a large-scale internal study found layout models produced better images across categories, and that its Reve 2.0 system is the strongest image generation model from a company valued below $1 trillion while using far fewer GPUs than some competitors.
The company presented reconstruction tests to support its claim. In those examples, images were rebuilt from layouts with increasing numbers of regions. Reve said more regions led to better fidelity, and that the method remained effective even without any pixels provided as input. It also said the same structure improves editing, since the layout can define exactly which elements should appear where.
Reve said the advantages of layout-based generation also appear when models scale up. The company claimed image quality improves as models grow larger and as they generate more regions, which it described as expanding the system’s visual reasoning context.
The larger message from the company is that image generation may eventually work more like program synthesis than simple prompting. In that view, people and AI systems would not just describe an image in words. They would read, write, and reason over a shared semantic representation that can be edited like code.
That vision is still emerging, but it reflects a notable change in how image models are being built. For users, the promise is clearer control. For developers, it suggests that the next frontier in generative imagery may depend less on clever wording and more on structured visual planning.