Digital art collections (e.g., WikiArt, Google Arts & Culture) have grown exponentially, yet automated analysis lags behind general object recognition. Art classification differs fundamentally from natural image classification: styles blend, artists imitate, and chronology matters. Existing datasets like Paintings91 (91 artists), WikiArt (over 1,000 artists but noisy labels), or OmniArt (large but uneven) suffer from label noise, class imbalance, or lack of temporal splits.
ArtClass v2 addresses these gaps:
Our contributions:
In the early ArtClass days, "bad anatomy" was sometimes passed off as style. In v2, artists are returning to fundamentals. The anatomy is solid, but the depiction is stylized. The lines are messy because the artist is confident, not because they are struggling. It’s the difference between a beginner scribbling and a master gesture drawing. artclass v2
Instead of one prompt, you can write a script:
[Prompt 1: "forest, golden hour, photorealism] -> [Prompt 2: apply watercolor filter strength 0.4] -> [Prompt 3: sharpen edges except on background trees]
This produces effects impossible with a single generation. Digital art collections (e