Reve (@reve) has released its image generation model "Reve 2.0," placing second in the Text-to-Image category of the community-voted benchmark Arena.ai. With OpenAI's GPT-Image lineup holding the top spot, Reve now stands as the highest-ranked independent foundation model lab.
June 3, 2026 · Reve
Reve 2.0 Lands at #2 in the Image Arena — Built by a 65-Person Lab
An independent lab's 4K "images-as-code" model ranks second overall on the Text-to-Image Arena right after launch — trained on roughly a tenth of the GPUs used by far larger competitors.
#2
Overall on the Text-to-Image Arena at launch
4K
Native 4096×4096 · 16 megapixels
~$0.01
Per image — $0.0067 to $0.04 range
65
People in the lab behind the model
Text-to-Image Arena · top of the leaderboard
Snapshot taken right after launch — Arena Elo shifts over time.
3
gemini-3.1-flash-image (nano-banana-2)
"Images as code" — a two-stage pipeline
1 · PLANNING
Layout as a structured intermediate — position, size, element descriptions
→
2 · RENDERING
The 4K image is drawn from the editable, code-like structure
→
RE-EDIT
Re-prompt or reposition single elements without breaking the whole
Praised
Layout editing feels "close to real design work"
Removes bottlenecks in iterative tasks
Native 4K + low-cost API suits large-scale exploration
Cautioned
Can excel on one prompt yet break across successive edits
Consistency and complex prompts still falter
Arena is only a reference — text and product visuals need more validation
A small lab competing on frontier turf signals a narrowing gap between large and small players.
$10 buys 7,500 credits — Create / Edit / Remix run ~5–30 credits each. Long-term behavior under heavy load and free-tier limits remains untested.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…
