Reve 2.0 Review: The Layout-First Image Model That Hit #2 at Launch
Every image model since 2022 has worked the same way: you type a sentence, you cross your fingers, you regenerate. Reve 2.0, released June 3, 2026, breaks that loop. Instead of describing the whole scene in one prompt and praying the model places things where you meant, you hand it a layout — a structured map of what goes where — and edit individual pieces without nuking the rest.
That's the pitch. The interesting part is that it mostly works, and it landed near the top of the Text-to-Image Arena on day one.
At a Glance
| Spec | Reve 2.0 |
|---|---|
| Released | June 3, 2026 |
| Developer | Reve — independent lab, ~65 people |
| Access | Closed — app + API, no open weights |
| Resolution | Native 4K |
| Headline feature | Layout-first — an editable JSON scene map, not a single prompt |
| Modes | Text-to-image, editing, multi-reference, in-image text |
| API price | ~$0.0067 per image (base tier) |
| Launch ranking | #2 on the Text-to-Image Arena at launch (behind GPT Image 2), per early reporting |
A caveat on that ranking: arena positions shift week to week, and launch-day numbers are partly a marketing event. Treat "#2 at launch" as a real signal that the quality is competitive, not as a fixed fact about where it sits today.
What "Layout-First" Actually Means
Here's the difference in one example.
The old way: "A coffee shop menu board with three drinks listed on the left, prices on the right, a logo at the top, warm lighting." The model reads that, makes its best guess at arrangement, and you get whatever it decided. Want the logo bigger? Reroll the whole image and hope the rest survives.
Reve 2.0's way: the scene is a structured object. Each element — the logo, the drink list, the price column, the background — is a named entry with its own position, size, and local description. You're not writing one sentence that has to carry everything. You're describing a layout the way you'd describe a webpage to a designer, block by block.
The payoff is addressable editing. Because every element has its own slot, you can change the price column's text without touching the logo, or nudge the menu board left without re-rolling the lighting. Each piece is independently editable, like layers in a design file rather than a single baked pixel grid. For anyone who's lost a good generation because one fix forced a full reroll, this is the part that matters.
It's not magic. Complex scenes still drift, and the model can ignore a layout constraint when the prompt and the structure disagree. But the working model in your head shifts from "describe and pray" to "compose and adjust," and that's a genuinely different way to generate images.
Reve isn't the only model that went this direction in June 2026 — Ideogram 4.0 shipped a layout-first approach the same day, using bounding-box JSON. The two arriving together is the real story: image generation is moving from prompt-first to layout-first, and Reve is the closed, 4K-native take on it.
Native 4K and the Typography Pass
Two things back up the layout pitch.
First, resolution. Reve 2.0 renders native 4K — not a 1K generation run through an upscaler, but 4K out of the model. For layout work that's not a luxury, it's load-bearing: fine text and small UI elements need the pixels to stay legible, and a layout model that topped out at 1024px would defeat its own purpose.
Second, text. Reve runs a dedicated typography pass, a stage focused specifically on rendering readable, correctly-spelled text inside the image. Text rendering has been the weak spot of image models for years — the warped, half-hallucinated lettering everyone learned to crop out. A model built around layouts of menus, posters, and product cards has to get text right, and Reve treats it as its own problem rather than hoping the main generator handles it.
How good is the text in practice? Competitive, by early comparisons, though GPT Image 2 still leads the text-rendering benchmarks by most accounts. Reve's advantage isn't that it writes the cleanest single line of text — it's that the text lives in an editable slot you can fix when it's wrong.
Closed, Paid, and Priced Per Image
Reve 2.0 is closed. There's no weight download, no self-hosting, no local run. You use it through the Reve app or the API, and that's the whole menu. It's built by Reve, an independent lab of around 65 people — small, focused, not attached to a big-cloud parent.
Pricing through the API starts around $0.0067 per image at the base tier. That's cheap per generation, which suits the actual workflow: layout-first means more iterations, not fewer, because the whole point is adjusting pieces and re-rendering. A model you poke at twenty times per final image needs to cost fractions of a cent per poke, and this does.
On content, Reve runs a standard prohibited-content policy — no pornography, no deepfakes of real people, no IP infringement — with no relaxed or "spicy" mode. If filter strictness is your deciding factor, the open-weight-doesn't-mean-uncensored breakdown compares Reve's policy against Ideogram 4.0's in detail, and the result is counterintuitive.
Where It Fits
Use Reve 2.0 when the image has structure. Menus, posters, ad creative, product cards, app mockups, anything where elements have to land in specific places and text has to be readable and fixable. This is the job it was built for, and the addressable-editing workflow is a real advantage over describe-and-reroll models.
Skip it when you want a single expressive image. A moody portrait, a painterly landscape, a one-off illustration — there's no layout to compose, so the headline feature does nothing for you, and a pure aesthetics model gets you there with less ceremony.
The honest limitation: it's new. Community data on real-world quality, edge cases, and how often the layout constraints actually hold is still thin three weeks after launch. The #2 arena debut says the raw quality is there. Whether the layout workflow holds up across messy real projects is the question only a few more months of use will answer.
Bottom Line
Reve 2.0 is the most interesting structural idea in image generation right now. Turning a flat prompt into an editable, addressable layout fixes the most frustrating thing about the medium — that one small fix used to cost you the whole image. Native 4K and a dedicated typography pass make that idea usable for the design-adjacent work it's aimed at.
It's closed and API-only, so there's no self-hosting story, and it's purpose-built enough that you wouldn't reach for it to make a single pretty picture. For layout-heavy work, it's worth watching closely and testing on a real project. For the other half of the June 2026 layout-first story, the open-weight one, read the Ideogram 4.0 review next.