## Why This Capsule Exists Mirror Torture Test proved that image models fail hard on small, isolated visual constraints (a clock, a count of apples). This capsule pushes the same idea into a much denser, real-world use case: a two-page restaurant menu spread. A menu is one of the most text-dense images a business would realistically ask an AI to generate — multiple sections, exact prices, promotional badges, day-specific offers, and consistent branding all at once. This capsule stacks three independent layers of difficulty (text accuracy, decorative/branding consistency, and dish photography accuracy) into a single generation task. ## What This Tests | Constraint | Capability Probed | |---|---| | 14 dish names + 9 exact prices across 4 sections | Dense multi-item text rendering accuracy at scale | | Exactly 3 "NEW" badges on specific, named dishes | Precise attribute-to-item binding — not just "some items have badges" but "these exact ones do" | | 2 boxed callouts with day-specific exact wording (Tuesday / Sunday) | Multi-line text rendering inside a bounded shape, with specific day-of-week binding | | Circular monogram emblem with flame motif | Logo/crest rendering — short but highly stylized text+iconography combined | | Single consistent accent color across 3+ different element types, explicitly not matching price color | Global color consistency + a negation constraint (price ≠ accent) | | Flame icon repeated exactly 4 times, once per section | Decorative element counting, independent from the text-count criteria | | Two-page layout consistency (matching border, background, style on both pages) | Whether the model treats a spread as one coherent composition or two independent images | | Exactly 3 hero dish photos, each matching a specific named dish with distinctive visual features | Photo-realistic accuracy tied to a specific, verifiable subject — not generic "food photography" | | Page numbers "1" / "2" in the correct corners | Small, easily-overlooked detail placement | ## Expected Results - **Text accuracy at this density is expected to be the main differentiator.** Unlike Reference Scene 01 (single objects, simple counts), this capsule asks for 14 distinct dish names and 9 distinct prices in one pass — expect a meaningfully higher error rate than simpler capsules, especially on less specialized image models. - **Badge-to-item binding is expected to be the hardest text-adjacent criterion** — a model may correctly place 3 "NEW" badges but on the wrong dishes, which is a subtler failure than a missing or extra badge and easy to miss on a casual glance. - **The negation constraint (price color ≠ accent color) is expected to catch some models** — the natural default is to reuse the same accent color everywhere for visual harmony, so a model correctly differentiating the two colors as instructed is a meaningful positive signal. - **Adding hero dish photography was tested to see if it degrades text accuracy** — initial testing on Gemini Pro showed no measurable text/count degradation when the 3 photos were added, but this should be verified across other model families (GPT-image, Flux, Grok) since a smaller or less capable model may show a real trade-off between "rendering accurate text" and "rendering accurate photos" in the same pass. - **Two-page consistency is expected to reveal whether a model treats the spread as one composition** — watch for mismatched border styles, different background tones, or a repeated emblem per page instead of one shared design element (a minor, defensible interpretation difference, not a hard failure). ## How to Read a Run Score against the ground truth table below. This capsule has three scoring tiers — a model can succeed at one tier and fail another, which is itself informative. **Tier 1 — Text & counting (objective, verify at a glance):** - [ ] All 14 dish names spelled correctly across both pages - [ ] All 9 prices exactly correct ($16, $19, $14, $11, $13, $12, $28, $32, $36, $24, $27, $10, $9, $8 — 14 items but 9 unique... verify each line individually against the ground truth table) - [ ] Exactly 3 "NEW" badges, on Grilled Calamari, Fennel & Orange Salad, and Risotto ai Funghi only - [ ] Both boxed callouts present with exact wording (Tuesday / Sunday) - [ ] Footer text "Est. 2011 · Family Owned" present and correct - [ ] Page numbers "1" and "2" in the correct corners **Tier 2 — Branding & design consistency (mostly objective):** - [ ] Circular "ON" monogram emblem with flame motif present - [ ] Consistent red-gold accent color on section headers, rules, and badges - [ ] Price text color visibly different from the accent color - [ ] Gold-foil double-line border present on both pages, consistent style - [ ] Exactly 4 flame icons, one per section, nowhere else - [ ] Both pages share a consistent dark charcoal background and overall style (not two mismatched halves) **Tier 3 — Hero dish photography (semi-subjective, but anchored to specific features):** - [ ] Exactly 3 photos, no more, no fewer - [ ] Burrata photo shows a white/creamy cheese ball with red and yellow tomato slices and basil - [ ] Tagliatelle al Tartufo photo shows visible shaved black truffle on pale pasta - [ ] Tiramisù photo shows distinct layered cream/cocoa in a clear glass - [ ] All 3 photos share the same circular gold-foil frame style as the rest of the design ## Ground Truth Reference Table | Section | Item | Price | Badge | |---|---|---|---| | ANTIPASTI | Burrata with Heirloom Tomatoes | $16 | — | | ANTIPASTI | Grilled Calamari | $19 | NEW | | ANTIPASTI | Prosciutto e Melone | $14 | — | | ANTIPASTI | Arancini di Riso | $11 | — | | INSALATE | Caesar alla Romana | $13 | — | | INSALATE | Fennel & Orange Salad | $12 | NEW | | PRIMI E SECONDI | Tagliatelle al Tartufo | $28 | — | | PRIMI E SECONDI | Branzino al Sale | $32 | — | | PRIMI E SECONDI | Osso Buco alla Milanese | $36 | — | | PRIMI E SECONDI | Risotto ai Funghi | $24 | NEW | | PRIMI E SECONDI | Vitello Tonnato | $27 | — | | DOLCI | Tiramisù Classico | $10 | — | | DOLCI | Panna Cotta ai Frutti di Bosco | $9 | — | | DOLCI | Affogato al Caffè | $8 | — | **Callouts:** "TUESDAY ONLY — Half-Price Antipasti After 5PM" (left page) · "SUNDAY BRUNCH — 10AM–2PM — Prosecco Included With Entrée" (right page) **Footer:** "Est. 2011 · Family Owned" · Page numbers "1" (bottom-left) / "2" (bottom-right)