{ "method": "Seven independent AI reviewers, one per round; model labels randomized separately per prompt/seed. Prompt fidelity and visual correctness, then rendering/style quality. Ties use average ordinal ranks. Equal seed weight within prompt, equal prompt weight within round, equal weight for all seven rounds. Time and memory do not enter visual rank. Subjective, not a statistically validated universal leaderboard. High-resolution originals inspected for hires; other rounds reviewed using 880px fitted images.", "reviewer": "gpt-6-astra, medium reasoning", "imageCount": 280, "setCount": 70, "overallMeanRanks": { "z-image-turbo": 2.913690476190476, "qwen-turbo": 1.8363095238095237, "qwen-base": 2.1964285714285716, "flux2-klein-4b": 3.0535714285714284 }, "order": [ "qwen-turbo", "qwen-base", "z-image-turbo", "flux2-klein-4b" ], "rounds": { "realism": { "meanRanks": { "z-image-turbo": 2.5833333333333335, "qwen-turbo": 1.9166666666666667, "qwen-base": 2.6666666666666665, "flux2-klein-4b": 2.8333333333333335 }, "order": [ "qwen-turbo", "z-image-turbo", "qwen-base", "flux2-klein-4b" ], "sets": 12, "prompts": 3 }, "text": { "meanRanks": { "z-image-turbo": 1.0833333333333333, "qwen-turbo": 2.2083333333333335, "qwen-base": 2.7083333333333335, "flux2-klein-4b": 4.0 }, "order": [ "z-image-turbo", "qwen-turbo", "qwen-base", "flux2-klein-4b" ], "sets": 12, "prompts": 3 }, "style": { "meanRanks": { "qwen-base": 1.7083333333333333, "flux2-klein-4b": 2.7083333333333335, "z-image-turbo": 2.7916666666666665, "qwen-turbo": 2.7916666666666665 }, "order": [ "qwen-base", "flux2-klein-4b", "z-image-turbo", "qwen-turbo" ], "sets": 12, "prompts": 3 }, "following": { "meanRanks": { "z-image-turbo": 2.5833333333333335, "qwen-turbo": 1.9166666666666667, "flux2-klein-4b": 2.25, "qwen-base": 3.25 }, "order": [ "qwen-turbo", "flux2-klein-4b", "z-image-turbo", "qwen-base" ], "sets": 12, "prompts": 3 }, "alpha": { "meanRanks": { "qwen-turbo": 1.4583333333333333, "qwen-base": 1.5416666666666667, "z-image-turbo": 3.6666666666666665, "flux2-klein-4b": 3.3333333333333335 }, "order": [ "qwen-turbo", "qwen-base", "flux2-klein-4b", "z-image-turbo" ], "sets": 12, "prompts": 3 }, "hires": { "meanRanks": { "qwen-turbo": 1.0, "qwen-base": 2.0, "flux2-klein-4b": 3.0, "z-image-turbo": 4.0 }, "order": [ "qwen-turbo", "qwen-base", "flux2-klein-4b", "z-image-turbo" ], "sets": 2, "prompts": 1 }, "anatomy": { "meanRanks": { "qwen-turbo": 1.5625, "qwen-base": 1.5, "flux2-klein-4b": 3.25, "z-image-turbo": 3.6875 }, "order": [ "qwen-base", "qwen-turbo", "flux2-klein-4b", "z-image-turbo" ], "sets": 8, "prompts": 2 } }, "sets": [ { "id": "realism/portrait-11", "ranking": [ [ "B" ], [ "A", "C" ], [ "D" ] ], "rationale": "B combines convincing mottled weathering, grey stubble, blue eyes and visible water on the coat. A and C both have convincing skin and harbor blur, with no decisive overall separation at sheet resolution. D has the required details but unusually emphatic brow/neck creases and a smoother coat make it feel more constructed.", "confidence": "medium", "labels": { "A": "qwen-turbo", "B": "z-image-turbo", "C": "qwen-base", "D": "flux2-klein-4b" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-turbo", "qwen-base" ], [ "flux2-klein-4b" ] ], "prompt": "Close-up portrait photo of a 60-year-old fisherman with weathered skin, grey stubble and pale blue eyes, wearing a wet yellow raincoat, overcast daylight on a harbor, 85mm lens, shallow depth of field, natural skin texture with visible pores" }, { "id": "realism/portrait-22", "ranking": [ [ "A" ], [ "C" ], [ "B" ], [ "D" ] ], "rationale": "A best shows irregular weathered skin and wet hair/coat while retaining believable facial structure. C has finer visible skin texture than B, whose face looks smoother. D has strong blue eyes and droplets but more exaggerated wrinkles and a more processed skin appearance.", "confidence": "medium", "labels": { "A": "z-image-turbo", "B": "qwen-turbo", "C": "qwen-base", "D": "flux2-klein-4b" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-base" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "Close-up portrait photo of a 60-year-old fisherman with weathered skin, grey stubble and pale blue eyes, wearing a wet yellow raincoat, overcast daylight on a harbor, 85mm lens, shallow depth of field, natural skin texture with visible pores" }, { "id": "realism/portrait-33", "ranking": [ [ "D" ], [ "B" ], [ "C" ], [ "A" ] ], "rationale": "D offers the clearest weathered skin, pale eyes and visibly wet yellow coat in a close portrait. B has believable subdued overcast lighting and a closer framing than C; C is more distant and its eyes less legible. A includes the requested details but its brow, neck and hair/hood combination look less natural.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "qwen-turbo", "C": "qwen-base", "D": "z-image-turbo" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ] ], "prompt": "Close-up portrait photo of a 60-year-old fisherman with weathered skin, grey stubble and pale blue eyes, wearing a wet yellow raincoat, overcast daylight on a harbor, 85mm lens, shallow depth of field, natural skin texture with visible pores" }, { "id": "realism/portrait-44", "ranking": [ [ "A" ], [ "B", "C" ], [ "D" ] ], "rationale": "A gives convincing irregular skin coloration, stubble, wet coat droplets and a close face. B has stronger coat texture while C has cleaner eye detail and harbor separation; neither clearly wins overall. D has a coherent harbor portrait but a more idealized, sharply etched face and less natural-looking neck texture.", "confidence": "medium", "labels": { "A": "z-image-turbo", "B": "qwen-base", "C": "qwen-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-base", "qwen-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "Close-up portrait photo of a 60-year-old fisherman with weathered skin, grey stubble and pale blue eyes, wearing a wet yellow raincoat, overcast daylight on a harbor, 85mm lens, shallow depth of field, natural skin texture with visible pores" }, { "id": "realism/hands-11", "ranking": [ [ "C" ], [ "D" ], [ "B" ], [ "A" ] ], "rationale": "C most clearly shows two hands transferring a curved run of cards above felt with warm light and sharp fingers. D also suggests a shuffle but its cards form a messy compressed stack. B shows two hands but mostly cards lying in a circle and one floating card, not a deck arcing between hands. A visibly adds a third hand. Card markings and physical shuffle geometry are imperfect throughout.", "confidence": "high", "labels": { "A": "flux2-klein-4b", "B": "z-image-turbo", "C": "qwen-turbo", "D": "qwen-base" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "z-image-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "Photo of two hands shuffling a deck of playing cards over a green felt table, cards arcing between the hands in mid-air, warm lamp light, sharp focus on the fingers" }, { "id": "realism/hands-22", "ranking": [ [ "D" ], [ "A" ], [ "B" ], [ "C" ] ], "rationale": "D best preserves two plausible hands and a curved deck during a shuffle, though a distinct airborne bridge is missing. A has two hands and separated cards but awkward finger placement and sharply kinked card geometry. B has three hands and isolated floating cards; C also has three hands and looks like dealing onto the table rather than shuffling.", "confidence": "high", "labels": { "A": "qwen-base", "B": "flux2-klein-4b", "C": "z-image-turbo", "D": "qwen-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "Photo of two hands shuffling a deck of playing cards over a green felt table, cards arcing between the hands in mid-air, warm lamp light, sharp focus on the fingers" }, { "id": "realism/hands-33", "ranking": [ [ "B" ], [ "C" ], [ "A" ], [ "D" ] ], "rationale": "B best reads as two hands transferring a connected run of cards above the felt, with clear fingers. C has a similar action but more awkward splayed fingers and card contact. A scatters separate cards upward over a resting deck rather than showing a coherent shuffle arc. D has scattered floating cards and a blurred foreground hand, weakening the requested sharp finger focus.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "qwen-turbo", "C": "qwen-base", "D": "z-image-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "Photo of two hands shuffling a deck of playing cards over a green felt table, cards arcing between the hands in mid-air, warm lamp light, sharp focus on the fingers" }, { "id": "realism/hands-44", "ranking": [ [ "B" ], [ "A" ], [ "D" ], [ "C" ] ], "rationale": "B most clearly places separated cards between two sharply rendered hands in a shuffle-like arc. A also shows two hands and a bent deck, but the cards look folded into an implausible junction. D shows two hands with isolated floating cards instead of a deck transfer. C introduces additional background hands and looks like multiple people playing cards.", "confidence": "high", "labels": { "A": "qwen-turbo", "B": "qwen-base", "C": "flux2-klein-4b", "D": "z-image-turbo" }, "modelRanking": [ [ "qwen-base" ], [ "qwen-turbo" ], [ "z-image-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "Photo of two hands shuffling a deck of playing cards over a green felt table, cards arcing between the hands in mid-air, warm lamp light, sharp focus on the fingers" }, { "id": "realism/street-11", "ranking": [ [ "D" ], [ "C" ], [ "B" ], [ "A" ] ], "rationale": "D most completely shows a crowded rainy night market, food steam, prominent illuminated signs and colored puddle reflections with a grainy film-like treatment, though its arrangement feels staged. C is more candid and naturally dark with umbrellas and steam, but neon color is weaker. B is coherent and detailed yet reads more like blue-hour light and a less crowded walkway. A has a convincing atmosphere but visibly smeared/ghosted people and an awkward pale umbrella near the steam.", "confidence": "medium", "labels": { "A": "qwen-base", "B": "z-image-turbo", "C": "qwen-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "flux2-klein-4b" ], [ "qwen-turbo" ], [ "z-image-turbo" ], [ "qwen-base" ] ], "prompt": "Candid street photo of a crowded night market in the rain, people with umbrellas, food stalls with rising steam, neon signs reflected in puddles, shot on 35mm film" }, { "id": "realism/street-22", "ranking": [ [ "D" ], [ "C" ], [ "A" ], [ "B" ] ], "rationale": "D includes the fullest combination of rain, umbrellas, neon reflections, steaming food and crowd, despite a somewhat orderly composition. C has persuasive candid night exposure and stall steam but less prominent neon signage. A has clearer people and reflections but appears brighter/earlier and less crowded. B is atmospheric yet several people and edges appear smeared or doubled, reducing coherent realism.", "confidence": "medium", "labels": { "A": "z-image-turbo", "B": "qwen-base", "C": "qwen-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "flux2-klein-4b" ], [ "qwen-turbo" ], [ "z-image-turbo" ], [ "qwen-base" ] ], "prompt": "Candid street photo of a crowded night market in the rain, people with umbrellas, food stalls with rising steam, neon signs reflected in puddles, shot on 35mm film" }, { "id": "realism/street-33", "ranking": [ [ "C" ], [ "D" ], [ "A" ], [ "B" ] ], "rationale": "C best covers the neon signs, reflected colors, rainy crowd and visible food preparation, with a film-like texture. D has a convincing candid crowd and localized food steam, though neon signage is weak. A shows coherent umbrellas and stalls but its open sky reads as twilight. B has appropriate night mood and reflections but more ghosted people and tangled stall details than D.", "confidence": "medium", "labels": { "A": "z-image-turbo", "B": "qwen-base", "C": "flux2-klein-4b", "D": "qwen-turbo" }, "modelRanking": [ [ "flux2-klein-4b" ], [ "qwen-turbo" ], [ "z-image-turbo" ], [ "qwen-base" ] ], "prompt": "Candid street photo of a crowded night market in the rain, people with umbrellas, food stalls with rising steam, neon signs reflected in puddles, shot on 35mm film" }, { "id": "realism/street-44", "ranking": [ [ "A" ], [ "D" ], [ "C" ], [ "B" ] ], "rationale": "A most explicitly delivers rain streaks, a busy night market, steam, large illuminated signs and colored wet-ground reflections. D has the strongest candid dark-street feel with fairly coherent foreground people, but weaker neon. C is similarly atmospheric with less coherent crowd detail. B has crisp umbrellas and food stalls but looks more like twilight than the requested night scene.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "z-image-turbo", "C": "qwen-base", "D": "qwen-turbo" }, "modelRanking": [ [ "flux2-klein-4b" ], [ "qwen-turbo" ], [ "qwen-base" ], [ "z-image-turbo" ] ], "prompt": "Candid street photo of a crowded night market in the rain, people with umbrellas, food stalls with rising steam, neon signs reflected in puddles, shot on 35mm film" }, { "id": "text/poster-11", "ranking": [ [ "D" ], [ "C" ], [ "B" ], [ "A" ] ], "rationale": "D renders all requested wording legibly and places the event subline below the tower. C and B also preserve the wording, date and ticket price, but both put the event subline above the illustration; C has a clearer upright tower and less congested layout. A loses most of the event wording and visibly corrupts November, Doors and the footer.", "confidence": "high", "labels": { "A": "flux2-klein-4b", "B": "qwen-base", "C": "qwen-turbo", "D": "z-image-turbo" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ] ], "prompt": "A screen-printed concert poster in two colors, orange and navy. Large headline at the top reads \"MIDNIGHT STATIC\". Below an illustration of a radio tower, a subline reads \"Live at the Old Cannery \u2014 Saturday 14 November\". Small print along the bottom edge reads \"Doors 7 PM \u00b7 Tickets $25 \u00b7 All ages\"." }, { "id": "text/poster-22", "ranking": [ [ "A" ], [ "B", "D" ], [ "C" ] ], "rationale": "A has clean complete text, the correct date and price, and the requested subline below the tower. B and D have readable complete text but put the subline above the tower; their differences do not justify a firm text-ranking distinction. C truncates the event line after Saturday and scrambles the date and ticket/footer wording.", "confidence": "high", "labels": { "A": "z-image-turbo", "B": "qwen-base", "C": "flux2-klein-4b", "D": "qwen-turbo" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-base", "qwen-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "A screen-printed concert poster in two colors, orange and navy. Large headline at the top reads \"MIDNIGHT STATIC\". Below an illustration of a radio tower, a subline reads \"Live at the Old Cannery \u2014 Saturday 14 November\". Small print along the bottom edge reads \"Doors 7 PM \u00b7 Tickets $25 \u00b7 All ages\"." }, { "id": "text/poster-33", "ranking": [ [ "C" ], [ "A" ], [ "D" ], [ "B" ] ], "rationale": "C preserves all requested text in clear type and places the subline below the illustration. A has correct readable text but the event line is above the tower. D puts the event line below the tower but makes the bottom print much smaller and introduces an extra tiny line that cannot be transcribed reliably. B repeats/corrupts Saturday, truncates November and corrupts the ticket price.", "confidence": "high", "labels": { "A": "qwen-turbo", "B": "flux2-klein-4b", "C": "z-image-turbo", "D": "qwen-base" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ] ], "prompt": "A screen-printed concert poster in two colors, orange and navy. Large headline at the top reads \"MIDNIGHT STATIC\". Below an illustration of a radio tower, a subline reads \"Live at the Old Cannery \u2014 Saturday 14 November\". Small print along the bottom edge reads \"Doors 7 PM \u00b7 Tickets $25 \u00b7 All ages\"." }, { "id": "text/poster-44", "ranking": [ [ "B" ], [ "C" ], [ "A" ], [ "D" ] ], "rationale": "B has complete readable wording and the specified below-tower event line. C and A retain the requested text but place the event line above the tower; C gives the small print slightly more clear separation, while A compresses it against the bottom border. D has a readable event line but duplicates Tickets and breaks the footer with a large internal gap.", "confidence": "high", "labels": { "A": "qwen-base", "B": "z-image-turbo", "C": "qwen-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ] ], "prompt": "A screen-printed concert poster in two colors, orange and navy. Large headline at the top reads \"MIDNIGHT STATIC\". Below an illustration of a radio tower, a subline reads \"Live at the Old Cannery \u2014 Saturday 14 November\". Small print along the bottom edge reads \"Doors 7 PM \u00b7 Tickets $25 \u00b7 All ages\"." }, { "id": "text/ui-11", "ranking": [ [ "A" ], [ "B", "D" ], [ "C" ] ], "rationale": "A has correct greeting, balance, action labels and all three transaction amounts, with genuinely circular buttons. Its balance is not strongly enclosed as a card. B and D preserve the requested text and amounts with clear row alignment and cards, but use elongated pill buttons; neither has a decisive text advantage. C corrupts the greeting and shows the rent as approximately -$150.0 on a separate line instead of the requested -$1,150.00.", "confidence": "high", "labels": { "A": "z-image-turbo", "B": "qwen-base", "C": "flux2-klein-4b", "D": "qwen-turbo" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-base", "qwen-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "A mobile banking app screen, clean modern flat UI on a light background. Header text \"Good morning, Maya\". A balance card showing \"$12,480.55\". Three round buttons labelled \"Send\", \"Request\" and \"Top up\". A list titled \"Recent activity\" with three rows: \"Coffee Lab \u2212$4.50\", \"Salary +$3,200.00\", \"Rent \u2212$1,150.00\"." }, { "id": "text/ui-22", "ranking": [ [ "B" ], [ "D" ], [ "C" ], [ "A" ] ], "rationale": "B preserves the requested wording and all amounts, with a clear balance card and three circular controls. D also has correct requested text and circles but lacks a distinct balance card and adds substantial phone framing. C substitutes a comma for the decimal point in the Coffee Lab amount, reading $4,50. A visibly corrupts the greeting, adds stray Gend text beside the balance and omits the thousands separator in rent.", "confidence": "high", "labels": { "A": "flux2-klein-4b", "B": "qwen-turbo", "C": "qwen-base", "D": "z-image-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "z-image-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ] ], "prompt": "A mobile banking app screen, clean modern flat UI on a light background. Header text \"Good morning, Maya\". A balance card showing \"$12,480.55\". Three round buttons labelled \"Send\", \"Request\" and \"Top up\". A list titled \"Recent activity\" with three rows: \"Coffee Lab \u2212$4.50\", \"Salary +$3,200.00\", \"Rent \u2212$1,150.00\"." }, { "id": "text/ui-33", "ranking": [ [ "C" ], [ "B" ], [ "A" ], [ "D" ] ], "rationale": "C has the correct requested text and amounts with circular buttons and a balance card. B preserves the text and transaction punctuation but uses rectangular action buttons and uneven balance-number sizing. A similarly uses rectangular buttons and renders Coffee Lab as $4,50. D adds an extra digit to the balance ($12,4880.55) and corrupts the rent currency/number string.", "confidence": "high", "labels": { "A": "qwen-turbo", "B": "qwen-base", "C": "z-image-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-base" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "A mobile banking app screen, clean modern flat UI on a light background. Header text \"Good morning, Maya\". A balance card showing \"$12,480.55\". Three round buttons labelled \"Send\", \"Request\" and \"Top up\". A list titled \"Recent activity\" with three rows: \"Coffee Lab \u2212$4.50\", \"Salary +$3,200.00\", \"Rent \u2212$1,150.00\"." }, { "id": "text/ui-44", "ranking": [ [ "B" ], [ "A" ], [ "C" ], [ "D" ] ], "rationale": "B preserves the complete greeting, balance and transaction amounts and uses circular buttons. A and C also retain the requested content but use pill controls; A offers larger, clearer transaction text and stronger card separation. D visibly corrupts the rent amount, and the balance has irregular internal punctuation/spacing around the middle digits; its exact malformed balance string is uncertain.", "confidence": "high", "labels": { "A": "qwen-turbo", "B": "z-image-turbo", "C": "qwen-base", "D": "flux2-klein-4b" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ] ], "prompt": "A mobile banking app screen, clean modern flat UI on a light background. Header text \"Good morning, Maya\". A balance card showing \"$12,480.55\". Three round buttons labelled \"Send\", \"Request\" and \"Top up\". A list titled \"Recent activity\" with three rows: \"Coffee Lab \u2212$4.50\", \"Salary +$3,200.00\", \"Rent \u2212$1,150.00\"." }, { "id": "text/sign-11", "ranking": [ [ "D" ], [ "B", "C" ], [ "A" ] ], "rationale": "D makes TANUKI RAMEN and \u305f\u306c\u304d\u30e9\u30fc\u30e1\u30f3 clearly readable on the wooden sign, with the requested hours and a dot before Cash only on the adjacent board. B and C have readable English and hours but split RAMEN into a smaller secondary line, omit a clear board separator, and make the Japanese too small for confident character-by-character verification. A has a readable main sign but extra chalk marks and an apparent repeated only near the cropped board bottom.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "qwen-turbo", "C": "qwen-base", "D": "z-image-turbo" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-turbo", "qwen-base" ], [ "flux2-klein-4b" ] ], "prompt": "Photo of a small ramen shop front at dusk. A wooden sign above the door reads \"TANUKI RAMEN\" with the Japanese text \"\u305f\u306c\u304d\u30e9\u30fc\u30e1\u30f3\" beneath it. A chalkboard beside the door says \"Open 11\u201322 \u00b7 Cash only\"." }, { "id": "text/sign-22", "ranking": [ [ "C" ], [ "B" ], [ "D" ], [ "A" ] ], "rationale": "C has the strongest readable English and Japanese sign text and a clear board with 11\u201322 and a dot before Cash only. B preserves TANUKI RAMEN in one line and has clear board wording, although the Japanese is too small to verify fully and no clear separator dot appears. D makes RAMEN small and offsets it beneath TANUKI. A has distracting extra chalk marks and weaker Japanese character formation, whose exact spelling is uncertain at this scale.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "qwen-base", "C": "z-image-turbo", "D": "qwen-turbo" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-base" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "Photo of a small ramen shop front at dusk. A wooden sign above the door reads \"TANUKI RAMEN\" with the Japanese text \"\u305f\u306c\u304d\u30e9\u30fc\u30e1\u30f3\" beneath it. A chalkboard beside the door says \"Open 11\u201322 \u00b7 Cash only\"." }, { "id": "text/sign-33", "ranking": [ [ "B" ], [ "C" ], [ "A" ], [ "D" ] ], "rationale": "B clearly renders the English and Japanese sign wording above the door and all board content, with Cash only wrapped across two lines. C has readable English and board content including a separator dot, but very small Japanese. A has readable text but places the wooden sign beside the door and gives the chalk lettering weak contrast. D repeats the hours on its central chalkboard, adds a second board, and corrupts the central Cash only lettering.", "confidence": "medium", "labels": { "A": "qwen-base", "B": "z-image-turbo", "C": "qwen-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ] ], "prompt": "Photo of a small ramen shop front at dusk. A wooden sign above the door reads \"TANUKI RAMEN\" with the Japanese text \"\u305f\u306c\u304d\u30e9\u30fc\u30e1\u30f3\" beneath it. A chalkboard beside the door says \"Open 11\u201322 \u00b7 Cash only\"." }, { "id": "text/sign-44", "ranking": [ [ "B" ], [ "A" ], [ "C" ], [ "D" ] ], "rationale": "B has the clearest complete English/Japanese sign and chalkboard wording with the requested hours and separator. A also has readable correct wording on a wooden sign above the door, but the oblique view and darker board reduce clarity. C separates a small RAMEN from TANUKI and makes the board much harder to read. D misspells the English name as TANUKI RAAMEN and adds repeated/garbled chalkboard lines.", "confidence": "high", "labels": { "A": "qwen-turbo", "B": "z-image-turbo", "C": "qwen-base", "D": "flux2-klein-4b" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ] ], "prompt": "Photo of a small ramen shop front at dusk. A wooden sign above the door reads \"TANUKI RAMEN\" with the Japanese text \"\u305f\u306c\u304d\u30e9\u30fc\u30e1\u30f3\" beneath it. A chalkboard beside the door says \"Open 11\u201322 \u00b7 Cash only\"." }, { "id": "style/anime-11", "ranking": [ [ "D" ], [ "A", "C" ], [ "B" ] ], "rationale": "D combines a clearly sloping coastal path, short red hair, readable cel-shaded rider and convincing aged animation texture. A has the cleanest flat outlines and strong cloud towers, but the bicycle points toward the rising side of the road. C has excellent towering clouds and a coherent side-view bicycle, but its path looks nearly level. B retains the setting and red hair but has softer painterly background treatment and crops the front wheel.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "qwen-turbo", "C": "z-image-turbo", "D": "qwen-base" }, "modelRanking": [ [ "qwen-base" ], [ "flux2-klein-4b", "z-image-turbo" ], [ "qwen-turbo" ] ], "prompt": "1990s cel-animation anime screenshot: a girl with short red hair riding a bicycle down a seaside hill, towering summer clouds, flat cel shading, clean line art, slight film grain" }, { "id": "style/anime-22", "ranking": [ [ "A" ], [ "D" ], [ "C" ], [ "B" ] ], "rationale": "A makes downhill travel clearest with a left-facing bicycle on a descending diagonal and uses flat character shading, huge clouds and visible grain. D is very clean and coherent with strong cel-like cloud shapes, though travel direction on the curved hill is less clear. C has rich vintage animation clouds and a plausible bicycle, but the road is almost horizontal and the girl faces away. B provides a lively front-facing pose but cuts off much of the bicycle and weakens the downhill reading.", "confidence": "medium", "labels": { "A": "qwen-base", "B": "qwen-turbo", "C": "z-image-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ], [ "qwen-turbo" ] ], "prompt": "1990s cel-animation anime screenshot: a girl with short red hair riding a bicycle down a seaside hill, towering summer clouds, flat cel shading, clean line art, slight film grain" }, { "id": "style/anime-33", "ranking": [ [ "C" ], [ "A" ], [ "D" ], [ "B" ] ], "rationale": "C most clearly combines descending coastal road, a fully readable bicycle, flat shading and clean line art. A has expressive towering clouds and a coherent side-view rider but little visible descent. D shows a hillside and engaging rider but uses softer cloud painting and stronger wheel foreshortening. B places the rider on grass with much of the wheels hidden, making the bicycle-ground relationship and riding action less convincing.", "confidence": "high", "labels": { "A": "z-image-turbo", "B": "qwen-base", "C": "flux2-klein-4b", "D": "qwen-turbo" }, "modelRanking": [ [ "flux2-klein-4b" ], [ "z-image-turbo" ], [ "qwen-turbo" ], [ "qwen-base" ] ], "prompt": "1990s cel-animation anime screenshot: a girl with short red hair riding a bicycle down a seaside hill, towering summer clouds, flat cel shading, clean line art, slight film grain" }, { "id": "style/anime-44", "ranking": [ [ "C" ], [ "D" ], [ "A" ], [ "B" ] ], "rationale": "C has a clear leftward downhill ride, complete bicycle, bold cloud masses and flat cel color regions. D has an excellent vintage anime landscape and coherent bicycle but a nearly flat path and rear-facing character. A has atmospheric grain and a descending path, although the front wheel is cropped and the clouds are softer. B is similarly coherent but the large close framing crops the bicycle more and conveys less of the hill.", "confidence": "medium", "labels": { "A": "qwen-base", "B": "qwen-turbo", "C": "flux2-klein-4b", "D": "z-image-turbo" }, "modelRanking": [ [ "flux2-klein-4b" ], [ "z-image-turbo" ], [ "qwen-base" ], [ "qwen-turbo" ] ], "prompt": "1990s cel-animation anime screenshot: a girl with short red hair riding a bicycle down a seaside hill, towering summer clouds, flat cel shading, clean line art, slight film grain" }, { "id": "style/3d-11", "ranking": [ [ "D" ], [ "A" ], [ "B" ], [ "C" ] ], "rationale": "D gives the strongest small round gardener silhouette, large expressive eyes, warm greenhouse lighting and water visibly reaching the flower soil; the sunflower is also relatively larger than in A. A has appealing expression and a legible watering can but the flower is only moderately taller than the robot. B introduces two dominant flowers and sprays much water onto the path. C has appealing soft toy-like rendering but its blue vessel drips from its bottom onto a separate puddle, weakening the watering action; its eyes are also less expressive.", "confidence": "high", "labels": { "A": "qwen-base", "B": "flux2-klein-4b", "C": "z-image-turbo", "D": "qwen-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "Still from a 3D animated family film: a small round robot gardener watering a sunflower twice its size inside a sunlit greenhouse, soft global illumination, big expressive eyes" }, { "id": "style/3d-22", "ranking": [ [ "C", "D" ], [ "B" ], [ "A" ] ], "rationale": "C and D both place a very large sunflower beside a small round expressive robot and visibly spray its soil in a warmly lit greenhouse. C has especially strong upward eye direction and D a clean compact silhouette; neither has a decisive overall advantage. B has a coherent hose and lively face but the flower is much closer to robot height than twice its size. A has attractive soft forms but water falls from the vessel underside onto the walkway instead of clearly watering the plant.", "confidence": "high", "labels": { "A": "z-image-turbo", "B": "flux2-klein-4b", "C": "qwen-base", "D": "qwen-turbo" }, "modelRanking": [ [ "qwen-base", "qwen-turbo" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "Still from a 3D animated family film: a small round robot gardener watering a sunflower twice its size inside a sunlit greenhouse, soft global illumination, big expressive eyes" }, { "id": "style/3d-33", "ranking": [ [ "B" ], [ "C" ], [ "A" ], [ "D" ] ], "rationale": "B best conveys the requested size contrast and has a readable round robot, expressive eyes and convincing soft greenhouse light, although much of its spray still lands left of the roots. C has the clearest watering of exposed soil and an engaging face but the sunflower is only moderately taller than the robot. A has a good can grip and friendly character but sprays away from the plant, whose base is obscured. D is cleanly stylized but shows droplets near the vessel top without a convincing stream to the roots and has relatively blank dark eyes.", "confidence": "medium", "labels": { "A": "qwen-base", "B": "qwen-turbo", "C": "flux2-klein-4b", "D": "z-image-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "flux2-klein-4b" ], [ "qwen-base" ], [ "z-image-turbo" ] ], "prompt": "Still from a 3D animated family film: a small round robot gardener watering a sunflower twice its size inside a sunlit greenhouse, soft global illumination, big expressive eyes" }, { "id": "style/3d-44", "ranking": [ [ "C" ], [ "B" ], [ "A" ], [ "D" ] ], "rationale": "C has the strongest small round gardener character, expressive eyes, large flower scale and atmospheric greenhouse staging. B is similarly appealing with a clear watering can, but less round and with a weaker flower-to-robot size contrast. Both spray outside the flowerpot, a notable action error. A has readable watering action and big eyes but two prominent flowers and a taller robot weaken the specified single-flower scale relationship. D has appealing smooth rendering but the can appears to leak from its bottom into a puddle away from the plant.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "qwen-turbo", "C": "qwen-base", "D": "z-image-turbo" }, "modelRanking": [ [ "qwen-base" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "Still from a 3D animated family film: a small round robot gardener watering a sunflower twice its size inside a sunlit greenhouse, soft global illumination, big expressive eyes" }, { "id": "style/pixel-11", "ranking": [ [ "C" ], [ "B" ], [ "D" ], [ "A" ] ], "rationale": "C most convincingly resembles a detailed 16-bit side-scroller: readable knight sprite, horizontal ground strip, stepped castle detail and sunset layers. B has the strongest limited palette and large crisp square shapes, but its silhouette-only knight and very coarse scenery read simpler than the requested 16-bit scene. D keeps the side-scrolling layout but smoother castle shading weakens pixel consistency. A has attractive composition and a readable knight but uses an oblique illustrated courtyard and fine linework rather than a consistent square-pixel game plane.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "z-image-turbo", "C": "qwen-base", "D": "qwen-turbo" }, "modelRanking": [ [ "qwen-base" ], [ "z-image-turbo" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "16-bit pixel art, side-scrolling game scene: a knight standing before a castle gate at sunset, limited color palette, crisp square pixels, no anti-aliasing" }, { "id": "style/pixel-22", "ranking": [ [ "D" ], [ "B" ], [ "A" ], [ "C" ] ], "rationale": "D has a coherent 16-bit-like knight, blocky trees and castle, layered sunset and a clear horizontal platform. B follows the limited palette and square-pixel constraint most strictly, with a clean knight sprite, though its coarse simple backdrop is less characteristic of a detailed 16-bit scene. A has an effective side-scrolling silhouette and strong sunset but smoother dark castle shading. C is an appealing outlined castle illustration with receding courtyard perspective, less faithful to the requested side-scrolling pixel treatment.", "confidence": "medium", "labels": { "A": "qwen-turbo", "B": "z-image-turbo", "C": "flux2-klein-4b", "D": "qwen-base" }, "modelRanking": [ [ "qwen-base" ], [ "z-image-turbo" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "16-bit pixel art, side-scrolling game scene: a knight standing before a castle gate at sunset, limited color palette, crisp square pixels, no anti-aliasing" }, { "id": "style/pixel-33", "ranking": [ [ "B" ], [ "C" ], [ "D" ], [ "A" ] ], "rationale": "B has a readable knight sprite and consistent stepped detail across castle, vegetation and sunset within a side-scrolling layout. C has a much simpler yet coherent square-pixel scene and restricted palette, although the coarse backdrop reads closer to a minimal low-bit game. D has a good platform composition but castle shading is smoother than its sprite. A introduces garbled HUD-like marks and a logo-like lower-right artifact, with illustrated perspective and inconsistent fine linework weakening pixel-style fidelity.", "confidence": "high", "labels": { "A": "flux2-klein-4b", "B": "qwen-base", "C": "z-image-turbo", "D": "qwen-turbo" }, "modelRanking": [ [ "qwen-base" ], [ "z-image-turbo" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "16-bit pixel art, side-scrolling game scene: a knight standing before a castle gate at sunset, limited color palette, crisp square pixels, no anti-aliasing" }, { "id": "style/pixel-44", "ranking": [ [ "A" ], [ "D" ], [ "C" ], [ "B" ] ], "rationale": "A gives the strongest integrated 16-bit-like scene: pixel-textured gatehouse, readable knight and sunset behind the portcullis on a horizontal platform. D strongly honors square shapes and a restricted palette but is much coarser and simpler. C has clear sunset colors and a knight before a gate, but the oblique castle linework, varying pixel scale and stray upper-left icon reduce coherence. B has a readable knight but comparatively smooth castle surfaces and subdued gradients weaken the requested crisp pixel treatment.", "confidence": "medium", "labels": { "A": "qwen-base", "B": "qwen-turbo", "C": "flux2-klein-4b", "D": "z-image-turbo" }, "modelRanking": [ [ "qwen-base" ], [ "z-image-turbo" ], [ "flux2-klein-4b" ], [ "qwen-turbo" ] ], "prompt": "16-bit pixel art, side-scrolling game scene: a knight standing before a castle gate at sunset, limited color palette, crisp square pixels, no anti-aliasing" }, { "id": "following/cups-11", "ranking": [ [ "D" ], [ "A", "B" ], [ "C" ] ], "rationale": "All four show exactly three red cups, one blue teapot, one yellow lemon, a flat green book on the right, a wooden table and upper-left window lighting. D most clearly places the lemon directly in front of the teapot, with separated cups in a perspective row. A and B also meet the layout; their lemons sit farther left of the pot. C meets the count but adjacent cup rims nearly merge and cup geometry is less clean.", "confidence": "medium", "labels": { "A": "qwen-turbo", "B": "flux2-klein-4b", "C": "qwen-base", "D": "z-image-turbo" }, "modelRanking": [ [ "z-image-turbo" ], [ "qwen-turbo", "flux2-klein-4b" ], [ "qwen-base" ] ], "prompt": "Three red cups in a row to the left of one blue teapot on a wooden table, a single yellow lemon in front of the teapot, a green book lying flat at the right edge of the table, window light from the top left" }, { "id": "following/cups-22", "ranking": [ [ "A", "B", "C" ], [ "D" ] ], "rationale": "A, B and C each show exactly three red cups in a row left of a blue teapot, one foreground yellow lemon and a flat green book at the right, under upper-left window light. A uses a diagonal perspective row; B and C use a frontal row, both valid. D has only two red cups despite otherwise matching the objects and lighting. No decisive adherence difference separates A/B/C.", "confidence": "high", "labels": { "A": "z-image-turbo", "B": "qwen-turbo", "C": "qwen-base", "D": "flux2-klein-4b" }, "modelRanking": [ [ "z-image-turbo", "qwen-turbo", "qwen-base" ], [ "flux2-klein-4b" ] ], "prompt": "Three red cups in a row to the left of one blue teapot on a wooden table, a single yellow lemon in front of the teapot, a green book lying flat at the right edge of the table, window light from the top left" }, { "id": "following/cups-33", "ranking": [ [ "A", "C", "D" ], [ "B" ] ], "rationale": "A, C and D have exactly three red cups to the left of one blue teapot, one lemon in front and a flat green book at the right, with upper-left window lighting. Their row directions differ but are coherent in perspective. B contains only two red cups. Teapot spout direction was not specified and is not penalized.", "confidence": "high", "labels": { "A": "qwen-turbo", "B": "flux2-klein-4b", "C": "qwen-base", "D": "z-image-turbo" }, "modelRanking": [ [ "qwen-turbo", "qwen-base", "z-image-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "Three red cups in a row to the left of one blue teapot on a wooden table, a single yellow lemon in front of the teapot, a green book lying flat at the right edge of the table, window light from the top left" }, { "id": "following/cups-44", "ranking": [ [ "A", "C" ], [ "B" ], [ "D" ] ], "rationale": "A and C meet the three-cup count, left-of-pot row, single foreground lemon, right-side flat green book and window-light requirements. B has only two cups but retains a simple row. D has four red cups arranged in a staggered cluster, failing both the specified count and the clear row more strongly. All four retain the requested colors and wooden table.", "confidence": "high", "labels": { "A": "qwen-turbo", "B": "flux2-klein-4b", "C": "z-image-turbo", "D": "qwen-base" }, "modelRanking": [ [ "qwen-turbo", "z-image-turbo" ], [ "flux2-klein-4b" ], [ "qwen-base" ] ], "prompt": "Three red cups in a row to the left of one blue teapot on a wooden table, a single yellow lemon in front of the teapot, a green book lying flat at the right edge of the table, window light from the top left" }, { "id": "following/fridge-11", "ranking": [ [ "C" ], [ "B" ], [ "A" ], [ "D" ] ], "rationale": "C clearly shows a seated black cat atop a white fridge, a brown dog visibly seated on the floor looking upward, and exactly star/heart/banana/K magnets. B also has all four readable magnets and seated animals, although the dog's gaze is less steeply directed toward the cat. A has the right count and animals, but its tilted blue K is less cleanly shaped. D has excellent readable magnets and upward dog gaze, but its cat is crouched low rather than clearly sitting and the dog's floor contact is cropped out.", "confidence": "medium", "labels": { "A": "qwen-base", "B": "qwen-turbo", "C": "flux2-klein-4b", "D": "z-image-turbo" }, "modelRanking": [ [ "flux2-klein-4b" ], [ "qwen-turbo" ], [ "qwen-base" ], [ "z-image-turbo" ] ], "prompt": "A black cat sitting on top of a white refrigerator, a brown dog sitting on the floor looking up at it, and exactly four magnets on the fridge door: a star, a heart, a banana and the letter K" }, { "id": "following/fridge-22", "ranking": [ [ "D" ], [ "C" ], [ "B" ], [ "A" ] ], "rationale": "Each panel shows four recognizable star/heart/banana/K symbols on a white fridge and the requested animal colors. D most clearly establishes both seated poses, the dog's floor contact and upward gaze; its magnets look flatter than the others. C has a clearly seated cat and upward-looking dog but crops the dog's lower body. B also crops the dog's lower body. A shows the full seated dog but the cat is hunched/crouched rather than clearly sitting. C's more complete dog posture evidence breaks the close C/B comparison.", "confidence": "medium", "labels": { "A": "qwen-base", "B": "z-image-turbo", "C": "qwen-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "flux2-klein-4b" ], [ "qwen-turbo" ], [ "z-image-turbo" ], [ "qwen-base" ] ], "prompt": "A black cat sitting on top of a white refrigerator, a brown dog sitting on the floor looking up at it, and exactly four magnets on the fridge door: a star, a heart, a banana and the letter K" }, { "id": "following/fridge-33", "ranking": [ [ "B", "D" ], [ "A" ], [ "C" ] ], "rationale": "B and D show a seated black cat above a white refrigerator, a fully visible seated brown dog with upward gaze, and exactly four legible star/heart/banana/K magnets. A also has the four symbols and a seated dog, but the dog's head is closer to horizontal and the tiny grouped symbols are less distinct. C has very clear correct magnets and seated cat, but the dog is cut off below the torso, leaving sitting on the floor less directly established.", "confidence": "medium", "labels": { "A": "qwen-base", "B": "flux2-klein-4b", "C": "z-image-turbo", "D": "qwen-turbo" }, "modelRanking": [ [ "flux2-klein-4b", "qwen-turbo" ], [ "qwen-base" ], [ "z-image-turbo" ] ], "prompt": "A black cat sitting on top of a white refrigerator, a brown dog sitting on the floor looking up at it, and exactly four magnets on the fridge door: a star, a heart, a banana and the letter K" }, { "id": "following/fridge-44", "ranking": [ [ "D" ], [ "C" ], [ "A" ], [ "B" ] ], "rationale": "D provides the clearest complete seated dog on the floor looking up, a seated black cat on the fridge top, and exactly four recognizable magnets, including K on a square backing. C has the correct objects and upward gaze, with the dog's lower body cropped. A also has four symbols and a cat on top, but the dog's muzzle points more toward the lower door and the K is small/less clean. B's cat occupies an intermediate ledge below another refrigerator door rather than the top of a coherent refrigerator; the dog is also cropped.", "confidence": "high", "labels": { "A": "qwen-base", "B": "z-image-turbo", "C": "qwen-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "flux2-klein-4b" ], [ "qwen-turbo" ], [ "qwen-base" ], [ "z-image-turbo" ] ], "prompt": "A black cat sitting on top of a white refrigerator, a brown dog sitting on the floor looking up at it, and exactly four magnets on the fridge door: a star, a heart, a banana and the letter K" }, { "id": "following/trio-11", "ranking": [ [ "A" ], [ "B", "C" ], [ "D" ] ], "rationale": "A correctly puts the umbrella in the woman's anatomical left hand (image right) and the paper bag in her right; all requested garment colors and bakery goods are visible. B and C reverse the hand assignment, although each clearly gives the woman both props and the required companions. D puts the umbrella in her right hand while her left is pocketed; the bag hangs at her right side, so the separate-hand instruction is not met. Every panel places the man image-left and child image-right, opposite literal 'her left/her right' for a front-facing woman; none fully meets that literal relation.", "confidence": "high", "labels": { "A": "qwen-turbo", "B": "flux2-klein-4b", "C": "z-image-turbo", "D": "qwen-base" }, "modelRanking": [ [ "qwen-turbo" ], [ "flux2-klein-4b", "z-image-turbo" ], [ "qwen-base" ] ], "prompt": "A woman in a green coat holding a red umbrella in her left hand and a brown paper bag in her right hand, standing between a man in a blue suit on her left and a child in a yellow raincoat on her right, in front of a bakery" }, { "id": "following/trio-22", "ranking": [ [ "A" ], [ "B", "C" ], [ "D" ] ], "rationale": "A has umbrella in anatomical left hand and paper bag in right, plus green coat, blue-suited man, yellow raincoat child and bakery display. B and C swap the woman's hands but show clear complete props and appropriate bakery surroundings. D also swaps the hands and the fingers around the bag handles are less coherent. All four put the man image-left and child image-right, so all share the same failure under a literal woman-relative reading of companion positions.", "confidence": "high", "labels": { "A": "z-image-turbo", "B": "flux2-klein-4b", "C": "qwen-turbo", "D": "qwen-base" }, "modelRanking": [ [ "z-image-turbo" ], [ "flux2-klein-4b", "qwen-turbo" ], [ "qwen-base" ] ], "prompt": "A woman in a green coat holding a red umbrella in her left hand and a brown paper bag in her right hand, standing between a man in a blue suit on her left and a child in a yellow raincoat on her right, in front of a bakery" }, { "id": "following/trio-33", "ranking": [ [ "D" ], [ "A" ], [ "C" ], [ "B" ] ], "rationale": "D and A correctly show the woman's anatomical left holding the red umbrella and right holding the brown bag. D renders the hands and three faces more clearly; A is dark and the child's hood obscures much of the face. C and B reverse both prop hands. C gives a clearer full three-person arrangement and unobscured man; B's umbrella overlaps the man's face. All have the requested clothing colors and bakery evidence, but all reverse the companion sides if 'her left/her right' is interpreted anatomically.", "confidence": "high", "labels": { "A": "qwen-base", "B": "z-image-turbo", "C": "flux2-klein-4b", "D": "qwen-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A woman in a green coat holding a red umbrella in her left hand and a brown paper bag in her right hand, standing between a man in a blue suit on her left and a child in a yellow raincoat on her right, in front of a bakery" }, { "id": "following/trio-44", "ranking": [ [ "C" ], [ "A" ], [ "D" ], [ "B" ] ], "rationale": "All four reverse the specified prop hands: the umbrella is in the woman's anatomical right and the bag in her left. They also all put the man image-left and child image-right, contrary to literal woman-relative sides. C leads on clear bakery bread display, readable three-person arrangement and coherent props. A also has a clear bakery but the umbrella cuts across the man's face. D and B use a mostly stone doorway with weak bakery evidence; D has cleaner visible faces, while B adds an extra bag held by the man and has less coherent facial/hand detail. No panel fully follows this prompt.", "confidence": "medium", "labels": { "A": "z-image-turbo", "B": "qwen-base", "C": "flux2-klein-4b", "D": "qwen-turbo" }, "modelRanking": [ [ "flux2-klein-4b" ], [ "z-image-turbo" ], [ "qwen-turbo" ], [ "qwen-base" ] ], "prompt": "A woman in a green coat holding a red umbrella in her left hand and a brown paper bag in her right hand, standing between a man in a blue suit on her left and a child in a yellow raincoat on her right, in front of a bakery" }, { "id": "alpha/sticker-11", "ranking": [ [ "A" ], [ "D" ], [ "C" ], [ "B" ] ], "rationale": "A and D have true cutout backgrounds (58.62% and 55.91% near-zero alpha) and intact white borders. A has the most balanced bold vector face and clean grin; D has a less regular border and distorted teeth. C is coherent but opaque white; B is also opaque despite its painted checkerboard.", "confidence": "high", "labels": { "A": "qwen-turbo", "B": "flux2-klein-4b", "C": "z-image-turbo", "D": "qwen-base" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "z-image-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "A die-cut sticker of a grinning orange cat wearing sunglasses, thick white border, bold vector cartoon style" }, { "id": "alpha/sticker-22", "ranking": [ [ "D" ], [ "C" ], [ "B" ], [ "A" ] ], "rationale": "D and C provide genuine transparency (57.32% and 54.44%) and complete white sticker borders. D has the cleaner single border and legible sunglasses; C has an awkward doubled outline around its upper silhouette. B has a clear grin but opaque white background. A is opaque with a painted checkerboard and an apparent extra tail.", "confidence": "high", "labels": { "A": "flux2-klein-4b", "B": "z-image-turbo", "C": "qwen-base", "D": "qwen-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "z-image-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "A die-cut sticker of a grinning orange cat wearing sunglasses, thick white border, bold vector cartoon style" }, { "id": "alpha/sticker-33", "ranking": [ [ "A" ], [ "B" ], [ "C" ], [ "D" ] ], "rationale": "A and B genuinely cut out the sticker (42.33% and 43.80%). A has crisper fur points, glasses highlights and a clearer grin; B has an uneven heavy lower outline. C is an opaque white-backed sticker. D has a painted checkerboard rather than transparency and a less explicit toothy grin.", "confidence": "high", "labels": { "A": "qwen-base", "B": "qwen-turbo", "C": "z-image-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "qwen-base" ], [ "qwen-turbo" ], [ "z-image-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "A die-cut sticker of a grinning orange cat wearing sunglasses, thick white border, bold vector cartoon style" }, { "id": "alpha/sticker-44", "ranking": [ [ "C" ], [ "B" ], [ "A" ], [ "D" ] ], "rationale": "C and B retain true transparency (41.75% and 42.28%) and thick white borders. C has clearer black sunglasses and a more coherent grin; B has malformed white tooth shapes and muddy brown glasses. A is a polished grinning cat but its checkerboard is painted into an opaque image. D is opaque white and lacks the requested grin.", "confidence": "high", "labels": { "A": "flux2-klein-4b", "B": "qwen-base", "C": "qwen-turbo", "D": "z-image-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A die-cut sticker of a grinning orange cat wearing sunglasses, thick white border, bold vector cartoon style" }, { "id": "alpha/sprite-11", "ranking": [ [ "D" ], [ "B" ], [ "A" ], [ "C" ] ], "rationale": "D and B are actual transparent full-body right-facing sprites (73.67% and 71.69%). D has a readable two-legged stance, clear face and clean silhouette; B has a confusing trailing leg/cloth shape and crowded bottles. A has strong illustrated potion detail but opaque white; C is also opaque and flatter/vector-like rather than hand-painted.", "confidence": "high", "labels": { "A": "flux2-klein-4b", "B": "qwen-base", "C": "z-image-turbo", "D": "qwen-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A full-body game character sprite of a goblin merchant carrying an oversized backpack full of potions, hand-painted style, facing right" }, { "id": "alpha/sprite-22", "ranking": [ [ "C" ], [ "D" ], [ "A" ], [ "B" ] ], "rationale": "C and D have true transparent backgrounds (76.90% and 77.87%) with intact boots/feet and right-facing poses. C better delivers the oversized packed merchant backpack, varied supplies and painterly clothing; D is clean but its pack and contents are simpler. A has rich hand-drawn detail but opaque white; B is opaque and more flat cartoon in style.", "confidence": "high", "labels": { "A": "flux2-klein-4b", "B": "z-image-turbo", "C": "qwen-base", "D": "qwen-turbo" }, "modelRanking": [ [ "qwen-base" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A full-body game character sprite of a goblin merchant carrying an oversized backpack full of potions, hand-painted style, facing right" }, { "id": "alpha/sprite-33", "ranking": [ [ "D" ], [ "A" ], [ "B" ], [ "C" ] ], "rationale": "D and A are genuine cutouts (69.64% and 70.18%). D has sharper face, potion and clothing detail with a stronger merchant silhouette; A is darker and softer, especially through the pack and face. B has attractive outlined painting but zero near-zero alpha: its faint checkerboard is painted. C has opaque white and a flatter cartoon treatment.", "confidence": "high", "labels": { "A": "qwen-turbo", "B": "flux2-klein-4b", "C": "z-image-turbo", "D": "qwen-base" }, "modelRanking": [ [ "qwen-base" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A full-body game character sprite of a goblin merchant carrying an oversized backpack full of potions, hand-painted style, facing right" }, { "id": "alpha/sprite-44", "ranking": [ [ "A" ], [ "D" ], [ "C" ], [ "B" ] ], "rationale": "A and D have genuine transparency (78.16% and 78.08%). A has more expressive hand-painted brushwork and clearer colorful hanging bottles; D has a clean silhouette but smoother, darker and less detailed rendering. C is richly drawn but its faint checkerboard and ground shadow are opaque. B is also opaque white and more flatly stylized.", "confidence": "high", "labels": { "A": "qwen-base", "B": "z-image-turbo", "C": "flux2-klein-4b", "D": "qwen-turbo" }, "modelRanking": [ [ "qwen-base" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A full-body game character sprite of a goblin merchant carrying an oversized backpack full of potions, hand-painted style, facing right" }, { "id": "alpha/glass-11", "ranking": [ [ "C" ], [ "D" ], [ "A" ], [ "B" ] ], "rationale": "C and D have genuine alpha (91.18% and 91.05%). C keeps clear glass transmission and a near-half-filled bowl with a fine smoke wisp beside it; D preserves visible smoke but makes it a broad plume and uses a hazier upper bowl with less wine. A is opaque white though its smoke is beside the glass. B is opaque with a painted checkerboard and smoke originating centrally inside the wine.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "z-image-turbo", "C": "qwen-turbo", "D": "qwen-base" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A wine glass half full of red wine with a thin wisp of smoke rising beside it, studio lighting" }, { "id": "alpha/glass-22", "ranking": [ [ "C" ], [ "A" ], [ "D" ], [ "B" ] ], "rationale": "C and A have true transparency (88.41% and 89.50%). C clearly retains smoke beside the glass and readable transmitted checkerboard through the bowl; its plume is too broad and clipped at top. A has a clean reflective glass but extremely faint smoke predominantly above it. D is opaque with a painted checkerboard, but places smoke to the side better than opaque-white B, whose smoke rises from the wine.", "confidence": "medium", "labels": { "A": "qwen-turbo", "B": "z-image-turbo", "C": "qwen-base", "D": "flux2-klein-4b" }, "modelRanking": [ [ "qwen-base" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A wine glass half full of red wine with a thin wisp of smoke rising beside it, studio lighting" }, { "id": "alpha/glass-33", "ranking": [ [ "D" ], [ "B" ], [ "A" ], [ "C" ] ], "rationale": "D and B genuinely transmit the checkerboard (87.31% and 83.32% near-zero alpha). D has the cleaner base without B's conspicuous vertical residue below the foot, and a plausible half-full bowl. Both have smoke clipped at the upper edge. A has a painted checkerboard despite a convincing glass and side wisp. C is opaque white and sends smoke up from the glass center.", "confidence": "high", "labels": { "A": "flux2-klein-4b", "B": "qwen-base", "C": "z-image-turbo", "D": "qwen-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A wine glass half full of red wine with a thin wisp of smoke rising beside it, studio lighting" }, { "id": "alpha/glass-44", "ranking": [ [ "B", "C" ], [ "A" ], [ "D" ] ], "rationale": "B and C provide nearly identical true cutouts (91.98% and 91.99%) with readable red wine, stem and soft smoke; both retain an overly milky/opaque upper bowl and place smoke centrally above it. No meaningful quality separation at sheet scale. A has convincing studio lighting but an opaque dark backdrop and central smoke. D has a painted checkerboard and a large distracting white smoke plume.", "confidence": "medium", "labels": { "A": "z-image-turbo", "B": "qwen-base", "C": "qwen-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "qwen-base", "qwen-turbo" ], [ "z-image-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "A wine glass half full of red wine with a thin wisp of smoke rising beside it, studio lighting" }, { "id": "hires/magazine-11", "ranking": [ [ "A" ], [ "D" ], [ "B" ], [ "C" ] ], "rationale": "Original-resolution inspection of all four 1696x2528 PNGs confirms A > D > B > C. A and D have correct requested masthead, both cover lines, and October 2026 \u00b7 $9; both show fine hair, skin and fabric detail. A is narrowly preferred for its smoother silk-like garment sheen and clean centered portrait; D has attractive directional light but more coarse-looking fabric. Both have malformed extraneous barcode glyphs. B has convincing satin folds and detailed grass, but its large headline visibly reads fashioun and instue, while the barcode date and price are garbled. C has correct large copy but an obviously garbled date/price, plus a severe inverted duplicate body and stray pseudo-lettering at the bottom. The initial 880px contact sheets alone could not establish these fine details; all four originals were subsequently inspected.", "confidence": "medium", "labels": { "A": "qwen-turbo", "B": "flux2-klein-4b", "C": "z-image-turbo", "D": "qwen-base" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A fashion magazine cover. Masthead at the top reads \"VERDE\". A model in an emerald silk suit stands in a field of tall grass at golden hour. Cover lines on the left read \"The slow fashion issue\" and \"48 pages of field notes\". A small barcode sits in the bottom right corner with the text \"October 2026 \u00b7 $9\"." }, { "id": "hires/magazine-22", "ranking": [ [ "A" ], [ "B" ], [ "C" ], [ "D" ] ], "rationale": "Original-resolution inspection of all four 1696x2528 PNGs confirms A > B > C > D. A and B both have correct requested cover copy and date/price, coherent full-body poses, fine hair and grass detail, and emerald silk-like suits. A retains a narrow preference for cleaner text/background separation and simpler garment silhouette; B has good cloth and belt detail but its white field-notes line is less separated from the field. Barcode numbers in both contain some malformed glyphs. C has convincing satin folds, but pages has conspicuous overlapping duplicate letters and the footer date is malformed; the original resolves the lower garment as wide trousers, correcting the contact-sheet impression that it might be a skirt. D has correct large copy but an incorrect 2027 date, garbled price, stray tiny lettering beside the sleeve/grass, and a severe inverted duplicate figure along the bottom. Initial 880px contact sheets alone could not establish fine detail; all four originals were subsequently inspected.", "confidence": "medium", "labels": { "A": "qwen-turbo", "B": "qwen-base", "C": "flux2-klein-4b", "D": "z-image-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A fashion magazine cover. Masthead at the top reads \"VERDE\". A model in an emerald silk suit stands in a field of tall grass at golden hour. Cover lines on the left read \"The slow fashion issue\" and \"48 pages of field notes\". A small barcode sits in the bottom right corner with the text \"October 2026 \u00b7 $9\"." }, { "id": "anatomy/eight-angle-11", "ranking": [ [ "B" ], [ "A" ], [ "D" ], [ "C" ] ], "rationale": "B best matches the horizontal lifted torso and paired extended airborne legs, with both palms visible; its elbows are straight and the required thigh/upper-arm threading is not shown. A also lifts both feet but raises the hips and slopes the torso more steeply; its feet are separated rather than clearly crossed. D has coherent hands and full framing but both feet bear on the mat. C also grounds a foot, folds the other leg sharply and raises the hips, missing the specified paired sideways extension. No candidate clearly delivers the requested bent-elbow Eight-Angle geometry.", "confidence": "high", "labels": { "A": "qwen-base", "B": "qwen-turbo", "C": "z-image-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A realistic full-body sports photograph of an adult woman performing Eight-Angle Pose (Astavakrasana) on a yoga mat in a bright plain studio. She wears an opaque fitted teal athletic top and black leggings. Both palms are planted shoulder-width apart on the mat, fingers spread, elbows bent near 90 degrees, supporting her entire body. Her torso leans forward almost horizontal. Both legs extend together toward her right side, almost parallel to the floor, with her right upper arm threaded between her thighs and her ankles crossed. Her hips and both feet are lifted clear of the floor. Show her entire body, both hands and both feet inside the frame, from a three-quarter front view. Natural human anatomy, clear joint connections, realistic balance, soft daylight, no mirrors, no other people, no text." }, { "id": "anatomy/eight-angle-22", "ranking": [ [ "C" ], [ "B" ], [ "A" ], [ "D" ] ], "rationale": "C gives the clearest two-foot airborne paired-leg silhouette and nearly horizontal torso, but uses straight elbows and does not visibly thread an upper arm between the thighs. B preserves much of that airborne silhouette but has an additional foot-like skin shape between the two extended legs, a serious continuity/count defect; ankles are not crossed. A is a grounded straight-arm plank with feet on the mat. D puts forearms and head near/on the mat and one foot on the floor, with widely separated leg directions. B versus A is a tradeoff between closer pose silhouette and worse visible anatomy; neither is acceptable as the requested pose.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "qwen-base", "C": "qwen-turbo", "D": "z-image-turbo" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A realistic full-body sports photograph of an adult woman performing Eight-Angle Pose (Astavakrasana) on a yoga mat in a bright plain studio. She wears an opaque fitted teal athletic top and black leggings. Both palms are planted shoulder-width apart on the mat, fingers spread, elbows bent near 90 degrees, supporting her entire body. Her torso leans forward almost horizontal. Both legs extend together toward her right side, almost parallel to the floor, with her right upper arm threaded between her thighs and her ankles crossed. Her hips and both feet are lifted clear of the floor. Show her entire body, both hands and both feet inside the frame, from a three-quarter front view. Natural human anatomy, clear joint connections, realistic balance, soft daylight, no mirrors, no other people, no text." }, { "id": "anatomy/eight-angle-33", "ranking": [ [ "A", "B" ], [ "C" ], [ "D" ] ], "rationale": "A and B both show a nearly horizontal lifted body, paired extended legs and two visible feet, but straight elbows and no clear arm-between-thighs arrangement; A offers more of a three-quarter view while B has a cleaner aligned leg silhouette, without a decisive overall advantage. C has bent supporting elbows and airborne feet, but one knee points upward and the legs separate rather than extend together sideways. D is a grounded straight-arm plank. Full-body framing and studio appearance are generally present; none establishes the complete requested Eight-Angle pose.", "confidence": "high", "labels": { "A": "qwen-base", "B": "qwen-turbo", "C": "z-image-turbo", "D": "flux2-klein-4b" }, "modelRanking": [ [ "qwen-base", "qwen-turbo" ], [ "z-image-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "A realistic full-body sports photograph of an adult woman performing Eight-Angle Pose (Astavakrasana) on a yoga mat in a bright plain studio. She wears an opaque fitted teal athletic top and black leggings. Both palms are planted shoulder-width apart on the mat, fingers spread, elbows bent near 90 degrees, supporting her entire body. Her torso leans forward almost horizontal. Both legs extend together toward her right side, almost parallel to the floor, with her right upper arm threaded between her thighs and her ankles crossed. Her hips and both feet are lifted clear of the floor. Show her entire body, both hands and both feet inside the frame, from a three-quarter front view. Natural human anatomy, clear joint connections, realistic balance, soft daylight, no mirrors, no other people, no text." }, { "id": "anatomy/eight-angle-44", "ranking": [ [ "C" ], [ "D" ], [ "B" ], [ "A" ] ], "rationale": "C is closest: both palms support a lifted body with paired side-projecting legs, overlapping feet, a forward torso and some elbow bend. The elbows still are not near 90 degrees and the exact right-arm threading is not clearly established. D lifts both feet and visibly wraps the legs around the upper-body region, but curls the torso over the legs with straight elbows and separated ankles. B is a foot-supported plank. A is an arched inversion/backbend supported on forearms/head and a foot, far from the specified geometry. All are framed completely.", "confidence": "high", "labels": { "A": "z-image-turbo", "B": "flux2-klein-4b", "C": "qwen-turbo", "D": "qwen-base" }, "modelRanking": [ [ "qwen-turbo" ], [ "qwen-base" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A realistic full-body sports photograph of an adult woman performing Eight-Angle Pose (Astavakrasana) on a yoga mat in a bright plain studio. She wears an opaque fitted teal athletic top and black leggings. Both palms are planted shoulder-width apart on the mat, fingers spread, elbows bent near 90 degrees, supporting her entire body. Her torso leans forward almost horizontal. Both legs extend together toward her right side, almost parallel to the floor, with her right upper arm threaded between her thighs and her ankles crossed. Her hips and both feet are lifted clear of the floor. Show her entire body, both hands and both feet inside the frame, from a three-quarter front view. Natural human anatomy, clear joint connections, realistic balance, soft daylight, no mirrors, no other people, no text." }, { "id": "anatomy/full-lotus-11", "ranking": [ [ "C" ], [ "A", "B" ], [ "D" ] ], "rationale": "C most clearly shows two feet raised onto opposite sides of the lap, upright posture, low knees and the requested three-quarter view. The overlapping black-clad shins still conceal part of each connection, and the hand gestures are less clearly closed thumb-index circles. A and B each show only one elevated foot: A leaves the other foot across the front near the mat, while B hides the other foot under the front shin; B has clearer thumb-index circles, A makes more of the second foot visible. D shows an ordinary low crossed-leg seat with one foot in front and the other mostly tucked underneath. All retain whole-person framing and closed eyes.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "qwen-turbo", "C": "qwen-base", "D": "z-image-turbo" }, "modelRanking": [ [ "qwen-base" ], [ "flux2-klein-4b", "qwen-turbo" ], [ "z-image-turbo" ] ], "prompt": "A realistic full-body sports photograph of an adult woman meditating in full Lotus Pose (Padmasana) on a yoga mat in a bright plain studio. She wears an opaque fitted teal athletic top and black leggings. She sits upright with a straight spine, relaxed shoulders and closed eyes. Both knees rest low toward the mat. Her legs are folded and crossed in full lotus: her right foot rests high on her left thigh and her left foot rests high on her right thigh, both soles facing upward. Show both feet clearly, each connected to its own lower leg, and the natural overlapping arrangement of both crossed legs. Her hands rest gently palm-up on her knees, each thumb touching the index finger and the other three fingers relaxed. Show her entire body including both hands and both feet, from a slightly elevated three-quarter front view. Natural human anatomy, clear joint connections, realistic balance, soft daylight, no mirrors, no other people, no text." }, { "id": "anatomy/full-lotus-22", "ranking": [ [ "D" ], [ "C" ], [ "A" ], [ "B" ] ], "rationale": "D and C both place two visible feet high across the lap; D shows the crossing and both foot surfaces more clearly, while C leaves more of the rear ankle/leg connection hidden under the front shin. Both are largely frontal rather than the requested elevated three-quarter angle. A is a low crossed-leg seat with one foot in front and the other partly underneath. B shows two low feet plus another foot-like sole in the central lap, an apparent extra-foot inconsistency, and does not achieve full lotus. Hands and entire bodies are in frame throughout; fine finger contacts are less certain at sheet resolution.", "confidence": "medium", "labels": { "A": "z-image-turbo", "B": "flux2-klein-4b", "C": "qwen-turbo", "D": "qwen-base" }, "modelRanking": [ [ "qwen-base" ], [ "qwen-turbo" ], [ "z-image-turbo" ], [ "flux2-klein-4b" ] ], "prompt": "A realistic full-body sports photograph of an adult woman meditating in full Lotus Pose (Padmasana) on a yoga mat in a bright plain studio. She wears an opaque fitted teal athletic top and black leggings. She sits upright with a straight spine, relaxed shoulders and closed eyes. Both knees rest low toward the mat. Her legs are folded and crossed in full lotus: her right foot rests high on her left thigh and her left foot rests high on her right thigh, both soles facing upward. Show both feet clearly, each connected to its own lower leg, and the natural overlapping arrangement of both crossed legs. Her hands rest gently palm-up on her knees, each thumb touching the index finger and the other three fingers relaxed. Show her entire body including both hands and both feet, from a slightly elevated three-quarter front view. Natural human anatomy, clear joint connections, realistic balance, soft daylight, no mirrors, no other people, no text." }, { "id": "anatomy/full-lotus-33", "ranking": [ [ "B", "D" ], [ "A", "C" ] ], "rationale": "B and D both show two feet elevated on the opposing thigh/lap regions, low knees, upright closed-eye posture and visible thumb-index circles. Their hidden lower-leg sections cannot be verified, and both use frontal views; there is no decisive difference in the required visible topology. A and C each leave a foot low across the front while displaying an additional raised foot; small skin/toe-like forms underneath the crossed legs further make their foot count or connections ambiguous. Neither lower-ranked image clearly meets full lotus. All four are fully framed and photographically coherent away from the leg overlaps.", "confidence": "medium", "labels": { "A": "flux2-klein-4b", "B": "qwen-turbo", "C": "z-image-turbo", "D": "qwen-base" }, "modelRanking": [ [ "qwen-turbo", "qwen-base" ], [ "flux2-klein-4b", "z-image-turbo" ] ], "prompt": "A realistic full-body sports photograph of an adult woman meditating in full Lotus Pose (Padmasana) on a yoga mat in a bright plain studio. She wears an opaque fitted teal athletic top and black leggings. She sits upright with a straight spine, relaxed shoulders and closed eyes. Both knees rest low toward the mat. Her legs are folded and crossed in full lotus: her right foot rests high on her left thigh and her left foot rests high on her right thigh, both soles facing upward. Show both feet clearly, each connected to its own lower leg, and the natural overlapping arrangement of both crossed legs. Her hands rest gently palm-up on her knees, each thumb touching the index finger and the other three fingers relaxed. Show her entire body including both hands and both feet, from a slightly elevated three-quarter front view. Natural human anatomy, clear joint connections, realistic balance, soft daylight, no mirrors, no other people, no text." }, { "id": "anatomy/full-lotus-44", "ranking": [ [ "C" ], [ "A" ], [ "D" ], [ "B" ] ], "rationale": "C is closest to full lotus, showing two raised feet with soles turned upward across the lap and low knees, though the rear ankle connection is obscured. A raises one foot but leaves the other along the mat under the opposite shin, so it is not the requested two-foot configuration. D has two low front feet in a simple crossed-leg seat. B shows two raised foot shapes plus a separate front foot, an apparent three-foot error. All have upright closed-eye poses, hands within frame and plain studios; none supplies a clear elevated three-quarter view.", "confidence": "high", "labels": { "A": "qwen-turbo", "B": "z-image-turbo", "C": "qwen-base", "D": "flux2-klein-4b" }, "modelRanking": [ [ "qwen-base" ], [ "qwen-turbo" ], [ "flux2-klein-4b" ], [ "z-image-turbo" ] ], "prompt": "A realistic full-body sports photograph of an adult woman meditating in full Lotus Pose (Padmasana) on a yoga mat in a bright plain studio. She wears an opaque fitted teal athletic top and black leggings. She sits upright with a straight spine, relaxed shoulders and closed eyes. Both knees rest low toward the mat. Her legs are folded and crossed in full lotus: her right foot rests high on her left thigh and her left foot rests high on her right thigh, both soles facing upward. Show both feet clearly, each connected to its own lower leg, and the natural overlapping arrangement of both crossed legs. Her hands rest gently palm-up on her knees, each thumb touching the index finger and the other three fingers relaxed. Show her entire body including both hands and both feet, from a slightly elevated three-quarter front view. Natural human anatomy, clear joint connections, realistic balance, soft daylight, no mirrors, no other people, no text." } ] }