Head to head: Recraft vs Krea 3 Edit
Recraft vs Krea 3 Edit
By Ryan Merket · Published
A head-to-head across eight image tasks tests style control, exact lettering, scene details, and spatial relationships. The key question is which model sticks more closely to the brief when visual polish and instruction-following pull in different directions.
Krea 3 Edit takes all eight saved task verdicts, with no ties: **8–0**. Its aggregate score is **68.7**, against Recraft’s **42.5**. That makes this a decisive result on this task sample—not a statistical claim about either model overall. The clearest gap is prompt fidelity. Krea’s Bauhaus kiosk stays graphic rather than drifting into photorealism; its navy poster preserves all three requested lines and punctuation; and its neon dumpling poster keeps the headline, subline, and bullet footer inside the frame. Recraft’s alternatives were often visually appealing, but missed the requested style or exact text in these comparisons. Krea also wins on scene construction: it delivers the five specified subjects and interactions in “Midnight Repair Parade,” and the pawnshop’s half-raised shutter, shopkeeper, display, and dawn reflections. In the bookshelf task, its centered one-point perspective is more consistent. The object-binding result was closer in the judges’ descriptions—Recraft looked more photographic, while Krea better conveyed the matte cube—but the saved verdict still goes to Krea. **Final call: Krea 3 Edit is the clear pick for this saved sample, especially when the brief depends on exact text, a named visual style, or several prescribed details landing together. This assessment applies only to these eight tasks.**
Bauhaus Kiosk Corner
A compact city newspaper kiosk on a tram-lined corner rendered in strict Bauhaus style, with geometric red-yellow-blue panels, black grid lines, asymmetrical but disciplined layout, flat graphic shapes, minimal shading, and elegant sans-serif signage reading MORGEN ECHO; pedestrians reduced to stylized geometric figures, circular tram wheels echoing nearby round lamps, crisp poster-like composition, bright daylight, 16:9.
OpenAI: GPT-6 Luna: Model B more fully captures the strict Bauhaus brief through flat geometric forms, disciplined color blocking, and stylized pedestrians, while retaining the tram and readable MORGEN ECHO signage. Model A has the correct kiosk panels and signage, but its photorealistic streetscape and realistic figures diverge from the requested poster-like graphic style. (Second judge pass, order swapped — scores are the average of both: Model B strongly captures the requested flat, disciplined Bauhaus poster st | Anthropic: Claude Fable 5.1: Model A renders a photorealistic street with a…
Legible multi-line text
A minimalist event poster with three lines of crisp, correctly-spelled text stacked and centered: 'NIGHT MARKET' large on top, 'Fridays · 6–11pm' in the middle, 'Riverside Pier 4' at the bottom, on a deep navy background, clean sans-serif, subtle grain.
OpenAI: GPT-6 Luna: Model B closely matches the requested minimalist navy poster and renders all three centered lines with the specified wording and punctuation. Model A is legible, but its photographic sign setting departs from the poster style and its middle line omits the centered dot and uses different capitalization and punctuation. (Second judge pass, order swapped — scores are the average of both: Model B presents all three lines clearly and correctly on a deep navy, subtly textured background, with the requ | Anthropic: Claude Fable 5.1: Model B delivers exactly what was asked: a mini…
Neon Dumpling Poster
A rain-speckled bus-stop poster in a narrow night street in Daegu, designed as bold contemporary street typography with clean hierarchy and high legibility, showing a stylized steaming dumpling illustration and the exact headline text EASY READ TEST, the subline 7 SPICE DUMPLINGS, and the small footer text LANE 14 • OPEN 22:10, all spelled exactly, evenly aligned, sharp, and easy to read from a few meters away; wet pavement reflections, teal-and-amber neon glow, frontal composition, 16:9.
OpenAI: GPT-6 Luna: Model B keeps the exact headline, subline, and footer—including the bullet—within a clearly framed, legible poster, with a fitting steamed dumpling illustration and strong rainy neon street setting. Model A has appealing atmosphere and large type, but the headline and subline spill beyond the poster and its footer uses a period instead of the requested bullet. (Second judge pass, order swapped — scores are the average of both: Model B presents all requested text exactly within a clearly framed, | Anthropic: Claude Fable 5.1: Model B renders all three text lines exactly (in…
Perspective & scale
A photorealistic one-point-perspective shot down a long empty library aisle, shelves receding to a single vanishing point, books getting consistently smaller with distance and nothing warped or wrongly sized, warm overhead light, 16:9.
OpenAI: GPT-6 Luna: Model A has a convincing, warmly lit aisle, but its foreground shelves and intermittent end panels make the perspective feel less uniform. Model B offers a cleaner, more symmetrical one-point perspective, with the books and shelves receding consistently toward the centered vanishing point. (Second judge pass, order swapped — scores are the average of both: Model B more cleanly fulfills the centered one-point perspective, with consistent shelf and book scale recession, an empty aisle, and warm ov | Anthropic: Claude Fable 5.1: Model B delivers a cleaner, more rigorous one-p…
Midnight Repair Parade
A coherent urban side-street scene at blue hour with five interacting subjects arranged naturally: a bicycle courier holding a dented silver helmet, a street tailor kneeling to pin up the courier’s torn orange jacket, a child on a folding stool offering a flashlight, a sleepy brindle dog tangled in a string of triangular shop flags, and an elderly fruit vendor leaning from a doorway with a crate of tiny pears; balanced composition with clear spacing and believable eyelines, layered storefronts, hanging cables, puddles, subtle motion, cinematic realism, 16:9.
OpenAI: GPT-6 Luna: Model B includes all five subjects and their requested interactions, including the torn orange jacket, flashlight, tangled flags, and crate of pears. Model A has a more cinematic street atmosphere, but several key details and interactions are missing or changed. (Second judge pass, order swapped — scores are the average of both: Model B includes all five requested subjects and their key interactions, with the kneeling tailor, flashlight-offering child, flag-tangled dog, helmet, and pear vendor c | Anthropic: Claude Fable 5.1: Model B renders all five subjects with the corr…
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.
OpenAI: GPT-6 Luna: Model B more convincingly captures the flat color planes, bold outlines, stylized foam, and muted palette of ukiyo-e, while Model A mixes the print-like wave with a photorealistic modern boat and seascape. Model A has dramatic scale, but Model B’s cohesive woodblock composition adheres more closely to the prompt. (Second judge pass, order swapped — scores are the average of both: Model B more closely follows the ukiyo-e brief with flat muted color planes, bold outlines, and a stylized wave, thou | Anthropic: Claude Fable 5.1: Model A collages a Hokusai-style flat wave onto…
Attribute binding
A photorealistic tabletop scene: a matte green cube to the LEFT of a glossy red sphere, a small blue cylinder BEHIND both, and a single yellow rubber duck sitting ON TOP of the cube. Colors must stay on the correct objects. Soft studio light, 16:9.
OpenAI: GPT-6 Luna: Model A has the more convincing photorealistic tabletop presentation and clearly places the duck atop the green cube, with the red sphere to its right and blue cylinder behind. Model B better conveys the cube’s matte finish, but looks more like a clean CGI render and its objects feel more crowded together. (Second judge pass, order swapped — scores are the average of both: Model B follows the requested matte green cube, glossy red sphere, blue cylinder behind both, and duck on top with a clean s | Anthropic: Claude Fable 5.1: Model B nails the bindings: a genuinely matte g…
Pawnshop Window at Dawn
Photorealistic early-morning street scene outside a small pawnshop just after sunrise: a half-open rolling metal shutter catching warm grazing light, a walnut display cabinet visible through glass, a shopkeeper with realistic skin and stubble adjusting a navy wool coat on a mannequin, tarnished brass watch chains, chrome camera bodies, cracked leather straps, dusty velvet trays, and a concrete sidewalk still damp from overnight cleaning; physically convincing reflections, soft long shadows, accurate material response for metal, wood, glass, skin, and fabric, shallow atmospheric haze, 16:9.
OpenAI: GPT-6 Luna: Model B more directly depicts the requested pawnshop details, including the shopkeeper adjusting a navy coat on a mannequin, a wooden display cabinet, cameras, a partly raised shutter, and wet pavement catching the dawn light. Model A has appealing street depth and warm reflections, but its shopkeeper and display are less faithful to the specific scene. (Second judge pass, order swapped — scores are the average of both: Model B more directly captures the requested half-open shutter, shopkeeper a | Anthropic: Claude Fable 5.1: Model B directly delivers the prompt's key elem…
Matchup powered by OpenRouter.