Head to head: CogView vs Cosmos 3 Super

The scoreboard is messy in the most interesting way. CogView posts the slightly higher aggregate score, 57.6 to 57.4, but Cosmos 3 Super takes the match on the statistical verdict with 75% confidence and wins more tasks outright, 5 to 3. That is not a contradiction so much as a reminder that averages can hide where a model actually lands cleanly on prompt fidelity.

Cosmos 3 Super wins by being the more reliable brief-follower on the harder compositional asks. It was plainly better on spatial layout, where it actually put the bed, desk, rug, and lamp where the prompt asked. It also took clockwork patisserie close-up with much stronger watch-part reconstruction and commercial macro discipline, mid-pour noodle sprint with a more believable roller-skating rescue and sense of speed, named art style with a more authentic ukiyo-e treatment, and golden-hour rooftop tasting with a scene that finally read like the requested rooftop rather than a nice balcony. That is a strong set of wins because they hinge on structure, style control, and prompt adherence, not just surface polish.

CogView’s wins are narrower but real. It deserved exact counting because it actually produced seven cups, which is the whole assignment, and Cosmos 3 Super simply did not. It also took color-bound tea still life by including the full object set that Cosmos dropped, and it came out ahead on negation mostly because Cosmos failed even harder. But that last category is hardly a bragging point: both models blew the no-plants, no-artwork constraints badly, which undercuts any claim of disciplined control on exclusion-heavy prompts.

So the verdict is straightforward. If you care about exact enumerations and slightly steadier average performance, CogView has a case. But if you want the model that more often nails the scene the prompt actually described, Cosmos 3 Super is the better pick here. The confidence level is only a lean, and the aggregate scores are nearly identical, so this is not a blowout. Still, on the evidence from these eight tasks, Cosmos 3 Super gets the nod.

How they were tested

We ran 8 fresh image tasks, generated on the fly for this matchup so neither model could prepare in advance, and had gpt-5.4 score each one. To cancel position bias, every task was judged twice — once in each presentation order — and every number reported here, including the headline totals, is the average of both passes. CogView scored 57.6 to Cosmos 3 Super's 57.4.

1. Spatial layout

A clean isometric illustration of a bedroom: a bed against the LEFT wall, a round rug centered on the floor, a desk under the WINDOW on the back wall, and a floor lamp in the FRONT-RIGHT corner. Flat-vector style, consistent perspective.

Winner: Cosmos 3 Super — Model B adheres more closely to the requested spatial layout: the bed is clearly against the left wall, the desk sits under the back-wall window, the rug is centered, and the floor lamp is placed in the front-right corner. Model A is attractive, but the lamp is not in the front-right corner and the rug is only loosely centered, making Model B stronger on prompt adherence and overall composition. (Second judge pass, order swapped — scores are the average of both: Model B follows the requested spatial layout much more closely: the bed is against the left wall, the desk is under the back-wall window, the round rug is centered, and the floor lamp is placed in the front-right corner in a clean isometric flat-vector style. Model A is more painterly than flat-vector and misplaces the floor lamp away from the front-right corner, though it remains visually appealing and well composed.)

2. Clockwork patisserie close-up

Ultra-detailed macro scene, 16:9: a cutaway view of an experimental patisserie window displaying a pistachio mille-feuille beside a miniature espresso machine rebuilt from brass watch parts; render the flaky pastry layers, powdered sugar crystals, pistachio crumb, woven linen napkin, hairline scratches in steel, engraved gear teeth, tiny steam droplets, and fingerprints on glass with forensic crispness, in photorealistic commercial tabletop style with a locked-off composition and clean neutral light.

Winner: Cosmos 3 Super — Model B adheres more closely to the prompt by clearly presenting a cutaway patisserie-window scene with a miniature espresso machine rebuilt from visible brass watch parts, stronger forensic detail in the metalwork, and a more locked-off commercial tabletop composition. Model A is attractive and photorealistic, but the espresso machine lacks the specified clockwork reconstruction and the scene misses the window/glass-display cues and several requested micro-details. (Second judge pass, order swapped — scores are the average of both: Model B adheres more closely to the prompt with a convincing brass watch-part espresso machine, stronger cutaway/detail emphasis, visible linen, glass reflections, and sharper forensic tabletop rendering. Model A is attractive and has appealing steam and pastry filling, but it misses the watch-part complexity and window-display feel, and its composition is less aligned with the specified commercial locked-off macro setup.)

3. Exact counting

A clean overhead flat lay on white marble of EXACTLY seven identical espresso cups arranged in a single straight row, each clearly separate and individually countable, even soft lighting, 16:9.

Winner: CogView — Model A is much closer to the prompt by showing exactly seven cups in a single straight row on white marble with clear separations, though the cups are not perfectly identical and the framing is slightly tight at the edges. Model B has cleaner spacing and a stronger flat-lay look, but it fails the exact counting requirement with only five cups and also omits the espresso entirely. (Second judge pass, order swapped — scores are the average of both: Model A clearly matches the prompt with exactly seven identical espresso cups in a single straight row on white marble, each separate and countable. Model B is clean and well-made, but it only shows five cups, so it fails the exact-count requirement despite decent composition and quality.)

4. Mid-pour noodle sprint

Dynamic food illustration in a semi-realistic cinematic style, 16:9: a roller-skating ramen courier in a saffron windbreaker rockets through a rain-slick alley, captured mid-stride as he twists to save a lacquer-red bowl of miso noodles; broth arcs through the air, scallions and sesame seeds trailing behind, one chopstick slipping free, coat hem snapping, wheels spraying water, background stalls streaked with motion blur to sell speed while the courier and bowl stay tack-sharp.

Winner: Cosmos 3 Super — Model B matches the prompt more closely with a true roller-skating courier in a rain-slick alley, strong motion blur in the background, and a tack-sharp subject and bowl; Model A is striking but misses key details by turning the bowl into a wheeled vessel and not clearly showing broth arcing or a believable courier save. Model B also has stronger cinematic composition and speed, though its helmet is an extra detail not requested and the noodle spill is somewhat restrained. (Second judge pass, order swapped — scores are the average of both: Model B matches the prompt more faithfully with a convincing roller-skating courier in a saffron jacket, a lacquer-red bowl held mid-save, strong rain-slick alley motion blur, and a sharper cinematic sense of speed. Model A is visually appealing but deviates by turning the bowl into a wheeled object, weakening the mid-pour rescue moment and overall prompt fidelity.)

5. Color-bound tea still life

Editorial still life, 16:9: on a brushed black basalt café counter, arrange exactly six drink objects with strict attribute binding and no color crossover — a matte cobalt ceramic teapot, a translucent amber glass tumbler of iced oolong, a polished copper cezve, a pale mint enamel milk jug, a smoky violet crystal sugar bowl, and a chalk-white porcelain saucer holding a single blood-orange macaron; hyperreal product photography, three-quarter view, crisp lens, soft studio side light that keeps each material unmistakable and every color locked to its own object.

Winner: CogView — Model A includes all six required objects and gets most of the color binding right, with only minor misses such as the tumbler lacking visible ice and the milk jug reading more like a lidded pot; Model B is more polished photographically but omits the polished copper cezve and breaks the strict color binding with a pink macaron instead of blood-orange. Model A also better matches the requested object set and editorial still-life arrangement on a dark stone counter. (Second judge pass, order swapped — scores are the average of both: Model A adheres more closely to the six-object brief by including the polished copper cezve and keeping the overall editorial still-life arrangement clean and readable, though it misses the iced tumbler and uses a lidded vessel instead of a clear mint enamel milk jug. Model B has stronger realism and lighting, but it omits the cezve entirely and breaks the color-binding requirement more noticeably with a pink macaron instead of blood-orange.)

6. Named art style

A ukiyo-e woodblock print of a fishing boat riding a large cresting wave at dawn, faithful to the flat color planes, bold outlines, and stylized foam of the tradition, muted indigo and cream palette.

Winner: Cosmos 3 Super — Model B is more faithful to the ukiyo-e woodblock print brief, with flatter color planes, a muted indigo-and-cream palette, bold contouring, and stylized foam that closely matches the tradition. Model A is striking and well-composed, but its saturated dawn gradient, oversized sun, and more modern illustrative finish make it feel less authentic to the named style. (Second judge pass, order swapped — scores are the average of both: Model B better matches the requested ukiyo-e woodblock print look with flatter color planes, muted indigo-and-cream palette, bold outlines, and stylized foam, while also presenting a convincing fishing boat in a cresting wave at dawn. Model A has strong composition and appeal, but its saturated sunset-like palette and more modern illustrative finish stray from the specified traditional ukiyo-e treatment and muted colors.)

7. Negation

A cozy reading nook with an armchair, a stack of books, and a mug of tea by a window — with absolutely NO plants, NO lamps, and NO artwork or picture frames anywhere in the frame. Warm afternoon light, 16:9.

Winner: CogView — Model A better matches the requested reading nook elements and warm afternoon light, but it still violates the negation constraints by including multiple plants and a framed picture. Model B is more aesthetically polished, yet it clearly fails the prompt with plants, a lamp, and wall artwork all present in the frame. (Second judge pass, order swapped — scores are the average of both: Model B has a warmer, more polished reading-nook composition, but it clearly violates the prompt with multiple plants, a lamp, and framed artwork. Model A also violates the prompt with several visible plants and framed wall art, though it avoids the lamp; both fail the key negation constraints badly, so neither meaningfully wins overall.)

8. Golden-hour rooftop tasting

Painterly realism, 16:9: a rooftop aperitif tasting at golden hour above the riverfront of a fictional city, with a long table of apricot spritzes, charred peach tartlets, and olive-oil cakes glowing in low sun; warm rim light catches bottle edges and glass condensation, long shadows stretch across terracotta tiles, distant cranes and water shimmer in honey-colored haze, and the whole composition should feel unmistakably late-day, serene, and sun-drenched.

Winner: Cosmos 3 Super — Model B adheres more closely to the prompt with a true rooftop setting, long table on terracotta tiles, visible riverfront cranes, and a convincingly late-day serene golden-hour atmosphere. Model A is attractive and polished, but it feels more like a small balcony café scene, misses the long tasting-table abundance and olive-oil cakes, and is less specific to the requested rooftop aperitif composition. (Second judge pass, order swapped — scores are the average of both: Model B better matches the rooftop aperitif tasting brief with a longer table setting, terracotta tiles, visible cranes, riverfront shimmer, and convincing late-day rim light and shadows, though some food details and the partial figure are off-brief. Model A is cleaner and painterlier, but it misses the long-table rooftop feel, lacks the specified cranes and olive-oil cakes, and reads more like a small balcony café scene than the requested expansive rooftop tasting.)

See every prompt and the full side-by-side outputs in the interactive Head-to-Head.