Reve Image

Reve Image is a family of text-to-image generative models developed by Reve AI, Inc., a Palo Alto, California startup, whose current flagship, Reve 2.0 (released 3 June 2026), ranks #2 on the Artificial Analysis text-to-image leaderboard with an Elo of roughly 1281, behind only OpenAI's GPT Image 2.^[17][19][21] Reve AI first emerged from stealth in March 2025 with Reve Image 1.0, a model that had circulated for several days on social media under the codename "Halfmoon" and, on launch, took the top position on the Artificial Analysis text-to-image Image Arena, edging out Midjourney v6.1, Google's Imagen 3, Black Forest Labs' FLUX 1.1 [pro] and other leading systems in blind, head-to-head human preference voting.^[1][2][4] Reve 2.0 is built around a "layout-first" method that translates a text prompt into a structured, editable layout before rendering in native 4K, an approach the company frames as treating "image generation as program synthesis," and which it says was "trained on 10x fewer GPUs" than larger competitors.^[19][20] The models are offered through the browser playground at app.reve.com (originally preview.reve.art), with a free tier, paid Lite and Pro plans, and a beta API.^[2][5][19]

Infobox

Scroll sideways for more →

History and background

The company

Reve AI was founded in 2024 and is headquartered in the San Francisco Bay Area, with offices listed in Palo Alto, and employed roughly 50 people by mid-2026.^{[5][7][8][23]} The founding team comprises Christian Cantrell, Michaël Gharbi, Taesung Park, Jon Watte, and Mike Speiser.^[7][9] Several founders come from Adobe Research's generative-imaging group: Gharbi and Park were both research scientists at Adobe before starting the company.^[9][10][11] Gharbi completed his PhD at MIT's Computer Science and Artificial Intelligence Laboratory under Frédo Durand, after undergraduate study in applied mathematics at École Polytechnique, and his prior research focused on computational photography, neural denoising, and image synthesis.^[10][9] Park earned his PhD at UC Berkeley with Alexei Efros and is widely associated with the image-to-image translation literature (his MS at Stanford was advised by Vladlen Koltun and Sergey Levine), with research interests in image editing using generative models.^[11] Cantrell, named in company filings as a founder and Chief Product Officer, came to Reve from a long tenure at Adobe in product and creative tooling roles.^[7][9] Watte serves as Founder CTO, and Mike Speiser, of Sutter Hill Ventures, is listed as a co-founder on Reve's portfolio profile.^[7][12]

Stealth period and "Halfmoon"

Through late 2024 and early 2025, an unidentified model under the cryptic identifier "Halfmoon" began appearing in the Artificial Analysis Image Arena and climbed to the top of the leaderboard before anyone publicly knew which lab was responsible.^[1][2][4] The arena is a blind A/B preference test: voters see two images generated from the same prompt and choose the better one, with the underlying model identities hidden until enough votes are recorded.^[4] On 25 March 2025 Reve unveiled itself as the team behind Halfmoon and renamed the production model Reve Image 1.0, releasing it as a free preview at preview.reve.art.^[1][2][3] The same week, secondary coverage in VentureBeat, Decrypt, and The Rundown documented the unusual circumstance of a previously unknown startup leading the global text-to-image leaderboard against well-funded incumbents.^[1][2][4]

Subsequent releases

On 15 September 2025 Reve announced "The New Reve," reorganising its product line around an image creator, a drag-and-drop direct-manipulation editor in beta, a creative-assistant chatbot, and a beta API.^[13] A core piece of that release was a new visual "layout representation" intended as an intermediate code between language and pixels, allowing region-level edits while preserving identity and spatial relationships, a concept that would later become the foundation of Reve 2.0.^[13] On 1 October 2025 the company released its first dedicated image-editing model, Reve V1, which debuted at #3 on the Artificial Analysis image editing leaderboard behind Google's Gemini 2.5 Flash ("Nano Banana") and ByteDance's Seedream 4.0.^[14][15] On 15 March 2026 Reve shipped Reve v1.5, advertised as operating in native pixel space at 4K resolution rather than relying on a compressed latent autoencoder, alongside new annotation tools ("Spotlight," "Draw," and "Select") and a redesigned single-panel interface.^[16] On 3 June 2026 the company released Reve 2.0, a native-4K model built on a new "layout-first" Large Layout Model, which debuted at #2 on the Artificial Analysis text-to-image leaderboard behind only OpenAI's GPT Image 2 (covered in detail in the Reve 2.0 section below).^[19][21]

Scroll sideways for more →

Reve 2.0

What is Reve 2.0?

Reve 2.0 is Reve AI's flagship text-to-image model, released on 3 June 2026, that generates native 4K images by first building an editable, structured "layout" of a scene and then rendering pixels from it.^[19][22] On its debut it entered the Artificial Analysis text-to-image leaderboard at #2, behind only OpenAI's GPT Image 2, with an Elo score of roughly 1280, about 125 points above the previous Reve v1.5.^[19][21] Artificial Analysis announced the result directly: "Reve 2.0 debuts at #2 on the Artificial Analysis Text to Image Leaderboard, behind only OpenAI's GPT Image 2."^[21] Reve described the model as "the best image generation model made by a sub-$1T company, trained on 10x fewer GPUs."^[19][20] The company, which employed roughly 50 people at the time, positioned the release as a return to the top of the board after its original Reve Image 1.0 had been overtaken by larger labs through 2025 and early 2026.^[17][23]

How does Reve 2.0's layout-first approach work?

Rather than mapping a text prompt straight to pixels, Reve 2.0 first converts the prompt into a "layout," a structured map of the objects, text, and regions in a scene, then renders the final image from that plan; the user can inspect and edit the layout both before and after rendering, so that moving a subject, rewriting a sign, or swapping a background causes the rest of the scene to adjust around the edit.^[19][20][22] Reve calls the underlying system a "Large Layout Model" that "takes any combination of layouts, instructions, and images as input, derives a layout from its internal thinking trace, and then renders the final pixels."^[20] The model is built by "continued pretraining and post-training of open-source large language models (thank you, Qwen team!) to learn spatial reasoning capabilities around our layout representation," trained via "a novel data pipeline on billions of images, bootstrapped from dense human annotations."^[20] Reve frames the overall bet as treating "image generation as program synthesis, so that humans and agents can read, write, and reason over a shared, code-like semantic intermediary," an "images as code" philosophy aimed at agent-driven creative workflows.^[19][20] In the company's ablation study, CLIP image similarity rose from 0.865 with no layout regions to 0.905 with 10 regions and 0.929 with 50 regions, which Reve cites as evidence that finer layout structure yields progressively more faithful reconstructions.^[20]

How does Reve 2.0 rank on the Artificial Analysis leaderboard?

As of July 2026, Reve 2.0 holds the #2 position on the Artificial Analysis text-to-image leaderboard with an Elo of about 1281, behind GPT Image 2 (Elo roughly 1339) and ahead of frontier systems including Google's Nano Banana 2 (Gemini 3.1 Flash Image) and Microsoft's MAI-Image-2.5.^[17][21] Reve emphasises that it reached this standing efficiently, stating publicly that it did so "with 1/10 to 1/100 resources of our competitors," a claim echoed by trade coverage that headlined the model's #2 finish as achieved "using 10x fewer GPUs."^[20][23]

How much does Reve 2.0 cost?

Reve offers Reve 2.0 through a free tier with a daily credit (or "energy") refresh, a Lite plan at about $7.99 per month, and a Pro plan at $19.99 per month, alongside a beta API priced at a fraction of a cent per image.^[22] Reviewers noted that this remained substantially cheaper than premium competitors, with Decrypt contrasting Reve's near-free API cost against Google's Nano Banana 2 (roughly 7 to 13 cents per image) and OpenAI's higher-priced tiers.^[22]

Model design and capabilities

A "logic-driven" generative system

Reve has not published a technical report, architecture diagram, or parameter count for Reve Image 1.0, and the company explicitly describes the model only as the result of a mission to "enhance visual generative models with logic," with marketing materials calling it a "logic-driven model" oriented toward prompt reasoning.^[1][3] The phrasing suggests heavier use of LLM-style language understanding to parse and decompose prompts, paired with a diffusion-style image decoder, although the precise division between the two remains proprietary.^[1][3] Independent commentary that has speculated about a "12-billion parameter hybrid" or specific module names should be treated as third-party guessing rather than confirmed specification.^[5][6] What Reve has documented publicly are the three design goals: prompt adherence, aesthetic quality, and typography.^[1][3] With Reve 2.0 the company was more forthcoming, disclosing that the model is a "Large Layout Model" derived by continued pretraining of open-source Qwen language models, a marked shift from the earlier silence on internals.^[20]

Prompt adherence

The most consistent observation in reviewer and arena commentary is that Reve Image 1.0 follows long, structured prompts unusually closely, faithfully placing multiple objects, attributes, and compositional cues that earlier diffusion models often dropped or merged.^[1][2][4] Reviewers attribute this in part to a prompt-interpretation step that decomposes a request before image synthesis begins, though the implementation has not been published.^[5] In Artificial Analysis's blind preference voting, this prompt adherence is one of the factors that drove its initial #1 ranking against Imagen 3, Midjourney v6.1, FLUX 1.1 [pro], and Recraft V3.^[1][4]

Typography and text rendering

Reve Image 1.0 was specifically engineered to render legible in-image text, a long-standing weakness of diffusion models that motivated competing systems such as Ideogram and Recraft V3.^[1][3] Editorial reviews in eWeek, Decrypt, and VentureBeat all single out the typography of Reve Image as one of its visible differentiators, with words, slogans, and posters appearing accurately spelled and stylistically coherent more often than in earlier open or proprietary models.^[1][2] The mechanism is again undisclosed; secondary sources describing a "proprietary typography engine trained on millions of font samples" are paraphrasing marketing copy rather than reporting verified internals.^[3][5]

Aspect ratios, resolution, and editing

The preview product offered configurable aspect ratios and a single base resolution at launch, with both natural-language editing of a generated image and uploads from disk supported through the chat interface.^[2][13] The 15 September 2025 platform update added a direct-manipulation editor, allowing users to draw on regions and instruct edits with text.^[13] Reve v1.5, in March 2026, raised native output to 4K and added annotation tools that let a user mark a target area before requesting an edit.^[16] Reve 2.0 generalised this into a full layout-first pipeline: the image is segmented and labelled into regions so a user can perform stable, surgical edits to a single element without regenerating the whole scene, while still rendering at native 4K.^[19][20][22]

Pricing and access

Reve Image launched as a browser-only product at preview.reve.art, requiring no payment to test.^[2][3] In the original preview, each account received 100 free credits upon signup and an additional 20 free generations per day, with paid top-ups priced at $5 for 500 generations (about $0.01 per image).^[2][5][6] Coverage at launch emphasised that this undercut Midjourney (subscription plans starting at roughly $8 per month and rising past $100 per month for heavy use) and Ideogram (similarly tiered subscriptions) by a wide margin on a per-image basis.^[2] As the product matured, Reve moved its main playground to app.reve.com and exposed a beta API for third-party developers.^[13] With Reve 2.0 in June 2026, the company kept a free tier alongside low-cost Lite and Pro subscriptions and a beta API priced at a fraction of a cent per image; see the Reve 2.0 section for current figures.^[19][22]

Performance and reception

Artificial Analysis leaderboard

The leaderboard maintained at artificialanalysis.ai/text-to-image ranks models by an Elo score derived from blind pairwise preference votes.^[4] At launch in March 2025, Reve Image 1.0 (under and after the "Halfmoon" label) held the #1 position, ahead of FLUX 1.1 [pro], Imagen 3, Midjourney v6.1, and Recraft V3.^[1][2][4] Reve Image's Elo at that point was reported by secondary sources as approximately 1167, though the live leaderboard recalculates continuously.^[17] In the months that followed, frontier systems including OpenAI's gpt-image-1, Google's Nano Banana family, and successive ByteDance Seedream releases shifted the top of the leaderboard, with the Reve V1 editing model in October 2025 debuting at #3 on the separate image-editing leaderboard behind Nano Banana and Seedream 4.0.^[15] By the spring of 2026, GPT Image 2, Nano Banana 2, and the FLUX.2 family occupy the top of the text-to-image board.^[17] That order changed again in June 2026, when Reve returned to the front with Reve 2.0, which debuted at #2 behind GPT Image 2 with an Elo around 1280 (about 125 points above Reve v1.5) and still held #2 with an Elo of roughly 1281 in July 2026.^[17][19][21]

Press coverage

VentureBeat described Reve Image 1.0 as "the new best AI image generation model" and emphasised the surprise of an unknown Palo Alto company beating Midjourney, Imagen 3, and Recraft V3 on a public leaderboard.^[1] Decrypt's coverage focused on the unusual combination of leaderboard performance with very low pricing, summarising the model as a "penny-per-image" competitor to Midjourney and Flux.^[2] eWeek highlighted readable in-image text as the breakout capability that distinguished Reve from existing tools.^[18] The Rundown AI's industry newsletter framed the launch as a "stealth startup dethroning image giants."^[4] The eponymous developer-aggregator post on pIXELsHAM also documented the 25 March 2025 launch, the Halfmoon codename, and the "ground-up" training claim.^[3] Coverage of Reve 2.0 in 2026 was similarly favourable: Decrypt reviewed it as "the best AI image generator for layout control," while AlphaSignal headlined its second-place finish as achieved "using 10x fewer GPUs."^[22][23]

Comparison with other text-to-image systems

Scroll sideways for more →

Compared with Midjourney, Reve emphasises text rendering and prompt adherence over a specific photographic house style; compared with Imagen 3, it is positioned as a standalone product rather than a feature of a broader assistant; compared with FLUX 1.1 [pro] and the broader Flux family, Reve does not currently distribute open weights and runs only as a hosted service.^[1][2][4] Against Ideogram and Recraft V3, the two prior text-rendering specialists, Reve's launch reviews specifically highlighted superior typography on long phrases.^[1][18] Where earlier systems, including Reve Image 1.0, expand a prompt internally before diffusion, Reve 2.0 exposes an explicit, editable layout as an intermediate representation, a design that distinguishes it from the prompt-to-pixels pipelines used by Midjourney, Imagen, and the Flux family.^[19][20]

Significance

Reve's debut illustrated three trends that have come to define the text-to-image market in 2025 and 2026.^[1][2][4] First, the field is no longer a duopoly of well-known web brands: a previously unknown Palo Alto startup beat better-funded incumbents on an open public benchmark in its first public week, much as ByteDance's Seedream and Google's Nano Banana later did from the large-platform side.^[1][4][15] Second, prompt adherence and in-image text rendering have become first-class quality axes, on par with raw photorealism, after years in which they were treated as secondary.^[1][2][18] Third, the per-image economics of high-quality image generation have collapsed: Reve's $0.01-per-image pricing at launch is roughly an order of magnitude cheaper than the per-image cost of running a Midjourney or Ideogram subscription at heavy use, putting price pressure on subscription-based incumbents.^[2] Reve's own trajectory adds a fourth theme: after Reve Image 1.0 was overtaken by larger labs through late 2025, the company returned to #2 on the text-to-image leaderboard in June 2026 with Reve 2.0, a comparatively small team reaching the frontier by betting on an editable "layout-first" representation rather than sheer scale, and claiming to have trained "on 10x fewer GPUs" than its rivals.^{[17][19][20][23]} That layout-first design also reframes image generation as an agent-friendly, "code-like" process, aligning image models with the broader move toward agentic and tool-using AI.^[19][20]

Limitations and criticisms

At launch, the most consistent reviewer complaints were the lack of any published technical paper or model card, the absence of an open API at preview time, no mobile application, and rather coarse moderation that pixelated outputs after the fact rather than refusing them up front.^[2] Because the architecture is undisclosed, reproducibility, third-party safety auditing, and academic comparison are constrained, a critique that applies to several proprietary image-generation systems.^[1][2] The natural-language editing flow in the preview was praised for simplicity but criticised for sometimes changing global aesthetics when the user only wanted a local edit, a problem the later direct-manipulation editor and v1.5 annotation tools are designed to address.^[2][13][16] Some reviewers in late 2025 and 2026 noted that, while Reve remained competitive on the editing and text-to-image leaderboards, Nano Banana, Seedream, gpt-image-1 and the FLUX.2 family had overtaken the original Reve Image 1.0 on raw quality at the top of the board by mid-2026.^[15][17] Reve 2.0 narrowed some of these gaps by disclosing its Large Layout Model design and shipping a beta API, but reviewers in mid-2026 reported that it can still drop elements of a prompt (so outputs need checking), that its identity preservation when editing from a human reference lagged behind Nano Banana 2 and ByteDance's Seedream 4.5, and that it sometimes ignored specific stylistic instructions.^[22]

Related work

• Midjourney and Midjourney V7, the long-standing subscription-based competitor.
• Imagen 3 and Imagen 4, Google's text-to-image models.
• Flux (text-to-image model), FLUX.1 and FLUX.2 from Black Forest Labs.
• Stable Diffusion and SDXL, the open-weight baselines.
• DALL-E 3 and gpt-image-1, OpenAI's image models.
• GPT Image 2, OpenAI's 2026 flagship image model, ranked just above Reve 2.0.
• Ideogram and Recraft V3, prior text-rendering specialists.
• Nano Banana, Google's Gemini-integrated image generation.
• Seedream, ByteDance's text-to-image system.
• Adobe Firefly, the incumbent commercial creative-tooling generator.
• Diffusion Transformer (DiT) and MMDiT, the architectural primitives commonly used by competing systems.
• Artificial Analysis, the benchmark hub that hosts the leaderboard on which Reve first attracted attention.

See also

• Text-to-Image Models
• AI Image Generation
• Diffusion model
• Latent diffusion model
• Prompt engineering for image generation
• Multimodal Models
• Generative AI

References

1. Carl Franzen, "The new best AI image generation model is here: say hello to Reve Image 1.0!", VentureBeat, 2025-03-26. https://venturebeat.com/ai/the-new-best-ai-image-generation-model-is-here-say-hello-to-reve-image-1-0. Accessed 2026-05-21. ↩
2. Jose Antonio Lanz, "New Reve Image Generator Beats AI Art Heavyweights MidJourney and Flux at a Penny Per Image", Decrypt, 2025-03-27. https://decrypt.co/311375/new-reve-image-generator-beats-ai-art-heavyweights-midjourney-and-flux-at-a-penny-per-image. Accessed 2026-05-21. ↩
3. pIXELsHAM, "Reve Image 1.0 Halfmoon: A new model trained from the ground up to excel at prompt adherence, aesthetics, and typography", pIXELsHAM.com, 2025-03-25. https://www.pixelsham.com/2025/03/25/reve-image-1-0-halfmoon-a-new-model-trained-from-the-ground-up-to-excel-at-prompt-adherence-aesthetics-and-typography/. Accessed 2026-05-21. ↩
4. The Rundown AI, "Stealth startup dethrones image giants", The Rundown AI newsletter, 2025-03-26. https://www.therundown.ai/p/stealth-startup-dethrones-image-giants. Accessed 2026-05-21. ↩
5. Reve AI, "Reve Image", official site preview.reve.art and app.reve.com, 2025-2026. https://preview.reve.art/. Accessed 2026-05-21. ↩
6. Magic Hour, "Reve Image 1.0 Review: The Next Midjourney/Flux?", magichour.ai blog, 2025. https://magichour.ai/blog/reve-image-review. Accessed 2026-05-21. ↩
7. Basis Set Ventures, "Reve, Portfolio Companies", basisset.com, 2025. https://www.basisset.com/portfolio-companies/reve. Accessed 2026-05-21. ↩
8. Crunchbase, "Reve, Company Profile and Funding", crunchbase.com, 2025. https://www.crunchbase.com/organization/reve-ai. Accessed 2026-05-21. ↩
9. Christian Cantrell, "Personal site", christiancantrell.com, 2025. https://christiancantrell.com/. Accessed 2026-05-21. ↩
10. Michaël Gharbi, "Personal site", mgharbi.com, 2025. https://mgharbi.com/. Accessed 2026-05-21. ↩
11. Taesung Park, "Personal site", taesung.me, 2025. https://taesung.me/. Accessed 2026-05-21. ↩
12. Jon Watte, "LinkedIn profile, Founder CTO at Reve AI", linkedin.com, 2025. https://www.linkedin.com/in/jonwatte/. Accessed 2026-05-21. ↩
13. Reve, "Introducing the New Reve", Reve blog, 2025-09-15. https://blog.reve.com/posts/the-new-reve/. Accessed 2026-05-21. ↩
14. Reve, "Reve's First Editing Model is One of the World's Best", Reve blog, 2025-10-01. https://blog.reve.com/posts/reve-editing-model/. Accessed 2026-05-21. ↩
15. Artificial Analysis, "Reve V1 debuts at #3 in the Artificial Analysis Image Editing Leaderboard", X (formerly Twitter), 2025-10-01. https://x.com/ArtificialAnlys/status/1973523022624743757. Accessed 2026-05-21. ↩
16. UI Things, "Reve Ships v1.5 Model, Annotation Tools, and a Redesigned Interface", uithings.com, 2026-03-15. https://uithings.com/reve-v1-5. Accessed 2026-05-21. ↩
17. Artificial Analysis, "Text to Image Leaderboard, Top AI Image Models", artificialanalysis.ai, 2026. https://artificialanalysis.ai/image/leaderboard/text-to-image. Accessed 2026-07-07. ↩
18. Fiona Jackson, "Reve AI's Image Generator Delivers What Midjourney Couldn't, Readable Text", eWeek, 2025-04-01. https://www.eweek.com/news/reve-image-ai-image-generator/. Accessed 2026-05-21. ↩
19. Reve, "Introducing Reve 2.0", Reve blog, 2026-06-03. https://blog.reve.com/posts/announcing-reve-2.0/. Accessed 2026-07-07. ↩
20. Reve, "The Layout Bet", Reve blog, 2026. https://blog.reve.com/posts/the-layout-bet/. Accessed 2026-07-07. ↩
21. Artificial Analysis, "Reve 2.0 debuts at #2 on the Artificial Analysis Text to Image Leaderboard, behind only OpenAI's GPT Image 2", X (formerly Twitter), 2026-06-03. https://x.com/ArtificialAnlys/status/2072443180004544909. Accessed 2026-07-07. ↩
22. Decrypt, "Reve 2.0 Review: The Best AI Image Generator for Layout Control", decrypt.co, 2026. https://decrypt.co/371033/reve-2-review-best-ai-image-generator-layout-control. Accessed 2026-07-07. ↩
23. AlphaSignal, "Reve AI's Reve 2.0 Hits #2 on Image Leaderboard Using 10x Fewer GPUs", alphasignal.ai, 2026. https://alphasignal.ai/news/reve-ai-s-reve-2-0-hits-2-on-image-leaderboard-using-10x-fewer-gpus. Accessed 2026-07-07. ↩