Controlling power matters more than discovering it.
Throughout history, the most transformative innovations were not simply those that unlocked new forces — but those that gave us mastery over them. Print did not merely copy text; it reshaped who could author reality. Flight did not merely lift bodies; it collapsed geography and accelerated every form of exchange that followed.
Generative AI is no different. The models that will define the next decade are not those that produce the most impressive single output — they are those that put precise, durable control in the hands of creators.
Reve 2.0 is built from that conviction.
Layout first. Image second.
Most image models collapse prompt and pixel into a single inseparable step. Reve separates them. A dedicated planning stage reasons about composition, spatial relationships, and visual hierarchy before a single pixel is rendered.
This separation means the model can reason about where things should go before deciding how they should look. The result is dramatically better composition — and a foundation that makes the rest of Reve's capabilities possible.
Planning is not a pre-processing step bolted onto an existing architecture. It is core to how Reve thinks.
If you can represent it, you can control it.
The dominant training paradigm pairs images with captions. Captions are lossy — they describe what a human found salient, not what the image actually contains. A caption cannot specify that the subject is 34% from the left edge, or that the shadows fall at a precise 17-degree angle.
Reve is trained on a code-based representation of images. Every spatial relationship, every element, every compositional decision is expressed in a structured, machine-readable form.
// Reve layout representation
{
"scene": {
"subject": { "x": 0.34, "y": 0.52, "scale": 1.0 },
"light": { "angle": 17, "intensity": 0.8 },
"depth": "shallow",
"mood": "cinematic"
}
}This foundation enables fine-grained control that caption-based approaches cannot match. What can be precisely represented can be precisely edited.
Read it. Edit it. Generate it.
Because Reve's images have a structured code representation, AI agents can do something no prior model enables: they can read an image as data, reason about its contents, modify the representation, and render a new image — all without any pixel-level manipulation.
An agent that understands Reve's format can answer questions about an image, make targeted changes, and maintain consistency across a multi-step creative workflow.
Original
Agent edit: shift light source
Agent edit: adjust depth
This is not prompt-based iteration. It is structured, deterministic editing — the same kind of control a developer has over source code.
Native 4K. No upscaling required.
Reve 2.0 renders natively at 4K × 4K — 16 megapixels per image. This is not a post-processing upscale. The model generates at full resolution from the start, preserving fine detail that upscaling algorithms inevitably reconstruct rather than recover.
For commercial workflows — print, advertising, editorial — resolution is not a luxury. It is a prerequisite. Reve eliminates the separate upscaling step that fragments most AI image pipelines.
Edit without accumulation.
Every time a traditional image model makes an edit, it re-processes what was already rendered. Artifacts compound. Fine details that survived the first generation degrade by the third. By the fifth iteration, the image may look substantially different from what was intended — not because of the edits, but because of the re-rendering.
Reve's code-based representation means editing happens at the structural layer, not the pixel layer. The rendered output is always derived fresh from a clean source — never re-processed.
Traditional workflow
Each edit re-renders from the previous output. Artifacts accumulate with every iteration. After 5+ edits, quality degrades measurably.
Reve workflow
Edits modify the structured representation. Each render is derived from the clean source. Quality is consistent across unlimited iterations.
Model and product, designed together.
Most AI image tools are built in two disconnected phases: a research team trains a model, then a product team figures out how to expose it. The result is a model that was never designed to be controlled — and a product that can only approximate control through prompting workarounds.
At Reve, the model and the editor are co-designed. The capabilities that make the editor powerful — structured layout, code-based representation, lossless editing — are not features added on top. They are properties of the model itself.
This is the only way to build a tool that gives creators genuine, durable control over what they make.
Cinematic. Grounded. Compositionally aware.
Reve 2.0 has a distinctive visual sensibility — informed by photojournalism, documentary photography, and the language of contemporary cinema. Not the antiseptic perfection of stock photography. Not the dreamlike distortion of early generative models.
Images that feel inhabited. Lit by real light. Composed with intention.
The model has a developed understanding of how the eye moves through an image, where tension lives, how negative space functions. This is not an accident of training data — it is a deliberate design goal.
Direct manipulation. Finally.
Because Reve's images are code, Reve's editor can offer something no image tool has before: direct manipulation of the underlying representation. Click a shadow. Move a subject. Adjust the depth of field. Each action modifies a structured parameter, not a freehand brushstroke over a rendered surface.
The editor is available at reve.com. It runs in the browser. It requires no local installation, no subscription to a cloud rendering pipeline, no waiting for a queue.
It is the fastest path from intention to image.