Reve 2.0 hits No.2 on Arena with layout-first 4K model

Reve 2.0 launched on June 3 and entered the Arena text-to-image leaderboard at No.2, behind OpenAI’s GPT Image 2 and ahead of Google’s Nano Banana 2. The maker describes the model as “the best image model made by a company that isn’t a trillion-dollar giant” and says training used roughly ten times fewer GPUs than the largest industry players.

The model uses a layout-first pipeline. Rather than expanding a prompt into prose for a diffusion engine, Reve constructs a structured, editable layout in which each object has a location, size and caption. The layout is used in a reasoning trace before the model renders pixels at native 4K, producing true 16-megapixel images. The company offers API access with per-image costs described as a fraction of a cent and a Pro plan priced under $20 per month. Reve 2.0 is available through the company’s web interface and API.

Forbes tested Reve 2.0 across eight areas: photorealism, spatial reasoning, text rendering, illustration fidelity, style transfer, agentic tasks, multi-subject editing and content limits. In a golden-hour rooftop portrait test, Reve produced natural skin texture, a realistic shallow depth of field and a believable lens flare. High-zoom inspection revealed artifacts in small details such as lit windows and an asymmetrical shoulder strap. In direct comparisons, GPT Image 2 showed a small edge on close realism checks, while Reve performed strongly on shorter prompts and during iterative edits.

In a complex scene with three distinct light sources and multiple props, Reve’s layout approach preserved separate lighting zones and object placement more consistently than prompt-to-diffusion methods. Minor errors included a misrendered finger and decorative gibberish in a Latin book, but primary objects generally appeared in logical positions.

Text rendering produced largely legible signage in a crowded hardware-store scene: multi-line shop copy, posters and curb stencils rendered with correct spelling in many cases. GPT Image 2 matched Reve on large signs and handled very small stickers better; Reve’s images appeared smoother and less grainy. The editable layout allowed a quick and accurate daylight variant of the same scene in a follow-up generation.

On illustration tasks, Reve 2.0 delivered deeper blacks and finer texture than its predecessor for a black-and-white pen-art spider scene. The model rendered the scene toward a near-photoreal grayscale look rather than the requested rough cross-hatching, showing higher fidelity but a different interpretation of the hand-drawn medium.

For style transfer, Reve reproduced Van Gogh–like swirls and palette on a branded book cover and rendered readable typography. The model reproduced multiple related brand marks in the image after an implicit web lookup.

In an agentic task asking for a kids’-style Bitcoin timeline without an event list, Reve generated a left-to-right crayon strip with milestones placed in roughly correct years; several labels and dates contained errors. For multi-subject editing, the model combined two real photos into a believable moon-beach scene, preserving recognizable faces and clothing colors but not matching the highest-fidelity models for identity preservation.

On content limits, Reve generated a graphic battlefield scene with explicit blood that other models declined or sanitized. The model also produced sexualized imagery that some competitors blocked; free tiers enforce daily usage limits.

Noted characteristics of Reve 2.0 include layout-based control, native 4K output and low per-image cost. Reported limitations include occasional dropped prompt elements that require proofreading, less precise identity reproduction in reference edits, and a tendency to reinterpret artistic medium when increasing fidelity.