Twelve recipes, rendered as written
Type the recipes as published — do you get the pictures they promise?
Unlike the other comparisons, the whole prompt changes here. These are the recipes from our fifteen curated examples, typed in unaltered.
The checkpoint differs by recipe. An anime model renders a photographic landscape as a painting, which would contradict the recipe text — the same landscape prompt was run through three checkpoints to pick. Each image page names the one used.
Three of the fifteen are missing. The school-uniform and chibi recipes fall outside what we publish, and pixel art was already measured as something these checkpoints do not do.
The cuts below share a seed and a sampler, but the prompt differs from cut to cut. Open any cut to see the full prompt it received.
What each cut shows
portrait photography- Photoreal portrait. The shallow depth of field and bokeh came through as written, but instead of
studio lightingthe scene became a forest in natural light. fantasy landscape- Floating islands, waterfall and aurora all made it in — a case where a long list of elements paid off.
cyberpunk cityscape- Neon and wet-asphalt reflections came in strong, though
flying carswas read as cars on the ground. watercolor illustration- White space, the autumn scene and a soft wash in the canopy came out as written, but the paper shows no grain and the figure's lines stay crisp. The quality prefix was left empty and only the recipe text used.
scifi concept art- Corridor perspective and glowing panels land in concept-art territory.
food photography- Rendered on the photoreal checkpoint, the plating and sauce read photographic.
architectural visualization- This one missed. The recipe asked for a
modern scandinavian villaand produced an empty interior with large windows rather than the building. Namingexteriorwould be the fix. oil painting style- The renaissance portrait framing, candles and gilded frame came through, but there is no impasto: the figure is outlined anime-style colouring.
dark fantasy character- The ash-grey battlefield and crimson accents came through, though the greatsword is not broken and the helmet shows no crack.
slice of life scene- Window, book and coffee cup all placed. Everyday scenes have few elements and reproduce steadily.
3d render portrait- Also missed.
3d render, stylized 3d animationproduced something close to a photograph rather than stylised 3D. That look needs a model or LoRA trained for it. nature photography- Fog, light shafts and water texture read photographic. This prompt was the yardstick in the checkpoint test.
How these were made
Shared prompt
Differs per cut — each cut page has it in full.
Model
Differs per cut — each cut page has it in full. · euler_ancestral/normal · steps 30 · CFG 5.5 · seed 777011
Takeaway
The recipes mostly work as written — but only when the checkpoint has seen that kind of picture. Feed a photographic recipe to an anime model and the model, not the prompt, decides the outcome.
How far this result reaches
Two cuts did not match their recipe and are published as they came out. Removing the misses would make the recipes look more reliable than they are.
Measured on a single Stable Diffusion family checkpoint. The same tags can behave differently on NovelAI, Midjourney or another checkpoint. This is a reference point, not a verdict.