Head to head: Bytedance Seedream V4.5 Edit vs Krea 3 Edit
Bytedance Seedream V4.5 Edit vs Krea 3 Edit
By Ryan Merket · Published
A head-to-head on whether image editing holds together under exact instructions: text, object relationships, perspective, and composition. The results also test how much polished lighting can compensate when framing or layout slips.
Krea 3 Edit wins where these edits are easiest to judge and hardest to fake: it follows the brief. It gets the bakery card’s four lines of wording right, places the duck on the green cube, and builds more convincing perspective in both the book aisle and attic tea-table scene. Seedream often looks polished, but in those tasks it breaks the text, misbinds the objects, or lets scale and geometry drift. The broader pattern is control, not raw prettiness. In the rainy-kitchen study, Krea keeps the requested objects visible and the scene coherent; Seedream’s warm-cool lighting and material detail are strong, but its tight crop cuts into the arrangement. The judges also tied reflections and exact counting: both models show seven cups, though Krea’s row is better framed, while neither fully nails the water reflections. Seedream’s clearest win is the Bauhaus desk palette, where its richer catalog-like treatment takes the task. That result matters, but it doesn’t outweigh Krea’s consistency across the more demanding layout and instruction-following tests. Krea takes five tasks to Seedream’s one, with two ties, and leads 65.0 to 57.4. **Final call: Krea 3 Edit is the clear winner, with a tentative lean.** Seedream can deliver a more atmospheric image, but Krea is the more dependable editor when the brief specifies what goes where—and exactly how it should read.
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 B presents a clear, symmetrical one-point-perspective aisle with books and shelves receding consistently toward a single vanishing point under warm overhead lights. Model A has the right setting and lighting, but its foreground shelving is oversized and visually cluttered, making the perspective and scale feel less consistent. (Second judge pass, order swapped — scores are the average of both: Model B delivers a clear, balanced one-point-perspective aisle with consistent scale recession, w | Anthropic: Claude Fable 5.1: Model A delivers a convincingly photorealistic,…
Bakery card with exact wording
A clean editorial product-style 16:9 photograph of a small countertop display in a neighborhood bakery, centered on a cream cardstock price sign clipped into a brushed brass holder beside a plate of three plum jam buns; the sign must be perfectly legible and read exactly, with line breaks preserved: "MIREN FOLD BAKERY" on the first line, "PLUM JAM BUNS" on the second line, "LOT 7" on the third line, and "$4.25" on the fourth line; use elegant condensed sans-serif typography in dark charcoal ink, straight-on composition, shallow depth of field, soft morning window light, and no extra text anywhere else in the image.
OpenAI: GPT-6 Luna: Model B presents the exact wording in the requested four clear lines, with a clean bakery display and three buns. Model A is warmly lit and polished, but its first line breaks incorrectly and the price is visibly malformed, undermining the exact-text requirement. (Second judge pass, order swapped — scores are the average of both: Model B closely matches the requested scene and renders all four lines of text correctly, with three buns and an elegant straight-on product composition. Model A is vis | Anthropic: Claude Fable 5.1: Model B renders the sign exactly as specified—f…
Reflections & glass
A photorealistic close-up of a chrome teapot on a polished dark marble counter next to a glass of water; the window and a red apple on the counter must be correctly reflected in BOTH the chrome and the water's surface, dramatic side light, 16:9.
OpenAI: GPT-6 Luna: Model A better delivers the requested close-up, dramatic lighting, and striking window and apple reflections across the chrome and glossy counter, though the water’s reflections are not fully clear. Model B is cleaner and more straightforwardly photorealistic, but its reflections in the water are also weak and the composition is less close-up. (Second judge pass, order swapped — scores are the average of both: Model B better matches the requested 16:9 close-up while keeping the teapot, glass, ap | Anthropic: Claude Fable 5.1: Model A delivers more dramatic lighting and viv…
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 B correctly places the yellow duck on the green cube, with the red sphere to its right and the blue cylinder behind both. Model A has attractive lighting but puts the duck on the red sphere and crops the objects heavily. (Second judge pass, order swapped — scores are the average of both: Model B clearly shows the matte green cube left of the glossy red sphere, with the blue cylinder behind them and the yellow duck sitting on the cube. Model A has polished rendering, but the duck is on the | Anthropic: Claude Fable 5.1: Model A places the duck on the red sphere instea…
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.
OpenAI: GPT-6 Luna: Model A shows seven cups in a straight row on marble, but the row is pressed low in the frame and the rightmost cup is cropped. Model B presents exactly seven clearly separate, identical cups in a tidy row with balanced spacing and a clean overhead look. (Second judge pass, order swapped — scores are the average of both: Model B presents exactly seven clearly countable, matching white espresso cups in a neat, centered row on white marble. Model A also shows seven cups, but the oversized empty sp | Anthropic: Claude Fable 5.1: Both Model A and Model B deliver exactly seven…
Bauhaus-only desk palette
A stylized stop-motion claymation still life of household desk objects arranged neatly on a matte tabletop in 16:9: a desk lamp, scissors, tape dispenser, stacked notebooks, mug, ruler, and pushpins, composed like a design catalog cover with crisp side lighting and soft shadows; restrict every visible color strictly to the named Bauhaus primary palette of red, blue, yellow, black, white, and neutral gray only, with no green, orange, purple, brown, beige, or other stray hues anywhere in the objects, shadows, highlights, or background.
OpenAI: GPT-6 Luna: Model B presents the requested desk objects in a clear, orderly catalog-style arrangement, while Model A crops several objects at the frame edges. Both use the Bauhaus colors effectively, though each has slight warm-toned details beyond the strict palette. (Second judge pass, order swapped — scores are the average of both: Model B presents the full, neatly arranged desk still life in the requested primary-and-neutral palette, with clear side lighting and a catalog-like composition. Model A uses | Anthropic: Claude Fable 5.1: Model A delivers a richer, more catalog-like cla…
Attic tea-table perspective
A photorealistic 16:9 view from waist height into a narrow attic breakfast nook under a slanted ceiling, with a long oak table running diagonally away toward a single clear vanishing point at the back-right window; on the near end sits a large cobalt-blue enamel teapot, beside it two medium stoneware bowls, then farther down three progressively smaller-looking apricots and a tiny brass hand bell near the window, all with correct relative scale and undistorted geometry; include mismatched wooden chairs tucked along both sides, exposed rafters, and floorboards whose lines converge consistently, morning light from the window, crisp depth cues, nothing warped or wrongly sized.
OpenAI: GPT-6 Luna: Model B has cleaner perspective, more balanced framing, and a coherent attic-table scene, though it omits the second bowl and places the window centrally rather than back-right. Model A includes both bowls and a back-right window, but its oversized foreground tableware and less consistent geometry weaken the composition. (Second judge pass, order swapped — scores are the average of both: Model B better preserves the requested sequence from the near teapot to the bowls, farther apricots, and bell | Anthropic: Claude Fable 5.1: Model B follows the object layout more faithful…
Rainy kitchen materials study
A highly photorealistic 16:9 close interior scene of a modest kitchen counter just after sunset during rain, lit by one warm under-cabinet strip and cool blue window spill: a brushed stainless steel moka pot with tiny water droplets, a nicked walnut cutting board, a half-peeled blood orange, a folded navy linen towel, a glazed ceramic bowl, and a human hand reaching in to set down a wet glass jar; emphasize physically believable reflections, soft contact shadows, subsurface skin tones, damp fabric, realistic wood grain, fingerprint-smudged metal, and the distinct material response of glass, ceramic, wood, steel, skin, and cloth.
OpenAI: GPT-6 Luna: Model A has striking warm-and-cool lighting and rich material detail, but its extreme crop leaves several elements cut off and makes the arrangement feel less cohesive. Model B presents the requested objects more clearly in a balanced, believable rainy-kitchen scene, with the hand naturally setting down the wet jar. (Second judge pass, order swapped — scores are the average of both: Model B presents the requested kitchen still life with all key objects clearly visible, balanced warm and cool lig | Anthropic: Claude Fable 5.1: Model A delivers the requested close-up materia…
Matchup powered by OpenRouter.