Testing Song-dynasty aesthetics in image models: local hits, unstable holistic judgment

sujingshen · x · 2026-08-18

The author documents a systematic test of whether image-generation models can reliably translate Song-dynasty aesthetics — not just producing "Chinese-style" images with azure greens, fine lines and silk texture, but making correct trade-offs when serving real articles, products and narratives.

Process: From material, coloring and line experiments to wireframe constraints, character narratives, knowledge covers and product posters; they tried color palettes, material cards, direction samples, structural wireframes, content routing and negative constraints, plus an A/B protocol on real articles, with some methods explicitly retired.
Conclusion: Models hit many local features but cannot yet reliably perform composite visual judgment — lines obeying objects, color attaching to form, negative space forming depth, main subjects answering the content, and coherent yet non-repetitive variation across a series.
Transferable method: Decompose aesthetics into judgeable tasks; manage proposition, structure, color, mood and counter-examples separately; change one variable at a time; validate with real use cases, not isolated pretty images; distill failures into asset boundaries and rules instead of ever-longer prompts.

Original post →

Testing Song-dynasty aesthetics in image models: local hits, unstable holistic judgment

More from Multimodal