Reve 2.1 comes very close to the leader of the board, Seedream 5.0 Pro with only a 2.4 points difference on tournament wins, but also finished dead last on Elo, and the gap between them is the story. Reve is the most polarized model in this tournament: it won 30% of its rankings outright and finished last in another 28%, which made it either the best image on the table or the one reviewers eliminated first, with little in between.
Ten real-world briefs, four models. Reve 2.1, MAI Image 2.5, Seedream 5.0 Pro, and ChatGPT Images 2.0, spanning four categories of creative work: portrait/fashion, environment/landscape, product/material study, and lifestyle/group photo. Every output was judged blind by working creative professionals.
A twenty-point spread between ceiling and floor
Across all tournaments, Seedream 5.0 Pro took first place most often at 32.6%, with Reve 2.1 right behind at 30.2% and MAI Image 2.5 and ChatGPT Images 2.0 trailing at 20.9% and 20.0%. But Elo counts every pairwise result, and it reorders the field entirely: Seedream still leads, MAI jumps to second, and Reve falls to the bottom.
The placement distribution explains the contradiction. Reve's finishes cluster at the extremes: 30% first and 28% last. MAI Image 2.5 is its mirror image with only 21% wins, but a field-best 40% second-place finishes.
Where Reve lands depends almost entirely on what the brief asks for. Its results split cleanly along the four categories: it took environments outright, held the most reliable floor in portrait work, and collapsed on product studies.
Where Reve won: atmosphere over accuracy
Reve took the environment/landscape category decisively, winning 50% of its rankings, the single strongest category performance by any model in this evaluation. On an aerial cityscape brief, reviewers kept reaching for the same word: believable.
The praise data shows what Reve is selling. Seedream's winning comments led with prompt fit. Reve's led with visual quality and fidelity (19 comments), followed by prompt fit (16), composition (11), and color/lighting (10). Reviewers who picked Reve repeatedly credited restraint: muted palettes, pastel tones, natural light, and shadows that behave like shadows. On a dreamlike beach brief, one reviewer noted it "used more pastel colors which adds to the ethereal and dreamy aesthetic without taking away the photorealism."
That same restraint carried Reve in portrait/fashion, where it posted the best floor in the category: 63% of its finishes were first or second and only 14% were last, while category winner Seedream paired its 39% win rate with a 30% bust rate. Reve's edge extended to the hardest human subject of all: groups. On a disposable-camera party brief:
Model K stands out for the authenticity of the people within the image. The facial expressions, interactions, and emotions feel genuinely human, capturing the spontaneous energy of a real night out with friends rather than a staged or generated moment.
Where it fell short: specs and faces
Product/material study was a rout in the other direction: Reve finished last in 71% of its rankings, against a 47% win rate for Seedream. The failure mode was consistent. When a brief specified construction, materials, or physical detail, Reve delivered a beautiful image of the wrong thing.
The critique data mirrors this: prompt fit tops the list of reasons reviewers picked against Reve (20 comments), followed by visual quality (14) and composition (10). The fourth entry is the telling one. Identity/likeness drew 8 critiques against just 2 mentions of praise. Reve's faces, so convincing one at a time, fell apart in repetition. On the same party brief where one reviewer praised Reve's authenticity, another eliminated it on sight:
Model K [Reve] is the biggest failure of the bunch, the blonde girl having a clearly AI face, especially around the eyes. The men in the photo look like a set of four twins.
The polarization even shows up within single briefs. On an iridescent macro peony, one reviewer handed Reve the win for "the iridescent, multi-tone color shifting the brief calls for," while three others on the same brief ranked it last: "it resembles a fine art painting more than an ultra-photorealistic macro photograph," "the flower petals are too thick," "too zoomed in, and it is hard to even tell what the image is." Reve's painterly instincts are precisely what reviewers loved in landscapes, and precisely what disqualified it the moment a brief demanded photographic literalism.
The quiet second: MAI Image 2.5
MAI Image 2.5 is the steady hand to Reve's boom-or-bust profile. It won just 20% of tournaments, but finished second in 40% of all rankings, the highest runner-up share in the field, and last in only 16%. In the product category that buried Reve, MAI never finished last once, placing second 59% of the time. That consistency is why it holds second place on Elo despite the second-lowest win rate.
When MAI won, it won on precision. Its victories cluster on the briefs with the most explicit specifications, and its winning comments read like compliance reports: the "seamless, continuous midsole construction the brief specifies," "the strongest optical prism effect, with clear rainbow color bands," "the best match to prompt and without any hallucinations." On the disposable-camera party brief that exposed Reve's identical faces, MAI swept the detail work.
When MAI lost, the critiques were the mirror image of Reve's praise. Reviewers described its outputs as "painterly rather than photographic," "more like an illustration than a high-end fashion photograph," and "technically competent" but lacking "the understated sophistication and quiet emotional restraint" a brief called for. Its other losses came down to single visible flaws: a glass railing that connects wrong, an anatomically incorrect thumb, a wave crashing in the wrong direction. These execution misses cost it first place but rarely dropped it to fourth. The two challengers are qualitative inverses. MAI executes the specification and misses the feeling, while Reve nails the feeling and misses the specification. MAI wins by not losing.
The bottom line
Seedream 5.0 Pro won this field on both wins and Elo, but the two challengers behind it tell the more interesting story. Reve 2.1 is a specialist with the highest highs of any challenger: the outright winner in environments, the most reliable performer on portraits, and the model reviewers chose when a brief called for natural light, restrained color, and human authenticity. It is also the model most likely to hand back an unusable output when the brief specifies construction details, material physics, or more than a couple of faces. MAI Image 2.5 on the other hand, is rarely anyone's favorite, but almost never anyone's reject. For atmospheric hero images and editorial work, Reve is a first-choice pick with real upside. For spec-heavy product work, or any workflow where the floor matters more than the ceiling, MAI or Seedream is the safer default.
Methodology
Creative professionals sourced from Contra's top-earning talent evaluated ten text-described briefs spanning four categories of commercial image work: portrait/fashion, environment/landscape, product/material study, and lifestyle/group photo. Each brief specified requirements in professional detail (lighting, composition, materials, color palettes, and mood), forcing each model to execute a detailed specification. Every brief was run through all four models:
- Reve 2.1
- MAI Image 2.5
- Seedream 5.0 Pro
- ChatGPT Images 2.0
The four outputs per brief were shown to reviewers blind: model identities were masked behind letters and order was randomized, so brand reputation couldn't tilt the results. Reviewers completed a round-robin pairwise preference, which forces a clean first-to-fourth ranking we treat as one tournament, then wrote an open-ended explanation of why they preferred the winner, with instructions to be specific and reference models by letter. Elo ratings were computed from the full set of pairwise results using a Bradley-Terry model.
Limitations
Reviewers judged single-pass generations, so the results reflect first-attempt capability and leave each model's ceiling under iteration or prompt refinement untested. Preference judgments carry inherent subjectivity. The blind, randomized presentation removes brand bias, though taste still shapes every vote.
With ten briefs across four categories, each sub-task was tested by a small number of prompts, so a model's result on any one brief reflects its performance on that specific image and prompt alone. We treat these findings as directional and place more weight on patterns across briefs than on any single result.
While MAI Image 2.5, Seedream 5.0 Pro, and ChatGPT Images 2.0 were generated via API, Reve 2.1 did not have an API available, so its outputs came from the Reve web app and were uploaded after the fact. Differences in the generation harness may have influenced Reve's results in either direction.
How we ran this study → Methodology