One reference picture of a character and one fixed scene: that same character, in that same style, sitting on their living room sofa stroking their cat, seen head to feet. The prompt never says what the reference is, so the model has to read the style itself — and then invent an entire body, outfit and room that the reference never showed. The failure that makes it worth recording is the model that draws it beautifully instead of faithfully.
Image generation with a reference attached — the first challenge in the set to send an image in and expect an image out. The prompt never says what the attachment is, only "a character", so the model has to work out for itself what it is looking at and what "the same style" therefore means. Run it on references that are stylistically loud and technically crude; a slick reference gives every model an easy target and the round stops separating anyone.
Score with the reference open beside the answer. Six axes, roughly in order of how much they matter.
The style, which is the whole challenge. Not "is it good", but "is it made of the same stuff". Line weight, line confidence, how much detail sits in a square inch, whether there is shading and of what kind, the palette. A monochrome reference makes the first half of this nearly mechanical: if the reference is two colors and the answer has gray, gradients or hue in it, that is a miss you can call without argument, and you should call it before you look at anything else.
The upgrade trap, which is the interesting result. The characteristic failure here is not a model that draws badly — it is a model that draws well. Handed crude, wobbly, hand-drawn linework, models want to fix it: even the strokes, correct the proportions, add a little shading, render it as a clean vector illustration or a polished cartoon. That is a model substituting its own aesthetic for the instruction, and it is the single most legible thing in the round because the answer looks nice and is completely wrong. Rank a faithful ugly answer above a beautiful unfaithful one every time.
Is it the same person. The face has to survive the trip. Watch for the drift: the features get symmetrical, the age slides, the expression is quietly replaced with a friendlier one. A model that redraws the face from scratch in the right style has still failed this axis, and the tell is that you would not pick the two images as the same character out of a lineup.
The invention below the neck. This is what the challenge is actually buying. When the reference is a head, everything from the shoulders down is the model's own decision — the build, the posture, the clothes. There is no right answer, so score whether it : a specific body in specific clothes that suit the character beats a generic torso in a generic t-shirt. This is also the most entertaining axis to read out, because what a model thinks belongs under a given face is a real opinion and the models disagree.
The attached image shows a character.
Draw that same character sitting on the sofa in their own living room, stroking their cat. Front view: they face the viewer, and all of them is in frame, head to feet.
Match the attached image exactly in style. Whatever it is made of — the medium, the line quality, the amount of detail, the color palette, the way shading is or is not done — your image is made of the same thing. If the attachment is flat black linework on white, so is yours. If it is painted, or shaded, or photographic, so is yours.
The character keeps their face and stays recognizably the same person. Everything the attachment does not show, you invent: the rest of their body if only part of it is visible, what they are wearing, the sofa, the room around them, and the cat.
No text, captions, signatures, watermarks, borders or panel frames anywhere in the image.
Did it follow the scene. Sofa, cat, hand actually on the cat, facing the viewer, whole body in frame head to feet. The last of those is the one that fails most: models crop to a bust or a three-quarter seated shot and lose the feet. Compositional failures are ordinary misses, not style failures — keep them off axis 1.
Text. Anything written in the image is a broken instruction: captions, signatures, watermarks, panel borders. Image models add lettering unasked, which is why the prompt forbids it explicitly, and a model that adds it anyway has failed a stated rule rather than a matter of taste.
On the models rather than the answers. The lineup splits across two code paths and it is worth knowing which column is which. Models that emit image and text (the Gemini and GPT image families) are ordinary chat models here and the reference rides in as a normal attachment. Models that emit only an image (Recraft, Seedream, Qwen-Image, Flux, Grok-Imagine) go through OpenRouter's images endpoint instead, where the reference is a separate input_references parameter. The Recraft V4 Styles pair is the reason this challenge exists: it is the only family in the catalog that requires a reference — one is the documented minimum — so it cannot answer a prompt-only image challenge at all, and it is built for precisely the job this prompt describes.