New-media art generation stops relying only on human ratings: CART rules steer diffusion sampling to FID 29.56
Synopsis
New-media art generation faces costly user-preference collection and diffusion sampling without interpretable guidance. The CART-rule-guided diffusion model first uses a conditional GAN to expand 2,000 real ratings into 10,000 virtual agents and 500,000 preference records, then converts CART decision-tree rules into a differentiable gradient injected into the reverse sampling of Stable Diffusion 1.5. Against five baselines it reaches FID 29.56, preference accuracy 0.734 and Hit Rate@3 0.512, improving over the strongest baseline RF-AIGC by 13.8%, 9.4% and 19.6%.
Interpretation
New-media art generation moves user preference from a post-generation filter into every step of diffusion sampling, so rules act during denoising. Earlier AIGC systems mostly used preference rules for post-hoc filtering, and preference models usually depended on small, costly human-rated datasets. On the test set the model reaches FID 29.56, preference accuracy 0.734 and Hit Rate@3 0.512, improving over the strongest baseline RF-AIGC by 13.8%, 9.4% and 19.6%, while IS and CLIP score also rise by 3.9% and 7.5%.
The preference signal comes from a virtual user pool synthesized by a conditional GAN rather than newly collected human ratings. Trained for 500 epochs on 2,000 real-user seed records, the conditional GAN produces 10,000 virtual agents and 500,000 preference records, with DKL 0.042, MMD 0.008 and a Fréchet-distance deviation of 6.3%, all below preset thresholds. A CART regression tree is trained on these synthetic preferences, its leaf rules are smoothed by a sigmoid into a differentiable violation measure, and the resulting gradient with strength 0.15 is injected every K steps into the reverse sampling of Stable Diffusion 1.5, whose backbone stays frozen.
A closed-loop feedback coordinator freezes the tree structure and adjusts only leaf-node weights and confidence values from each round's agent-scoring deviation. Removing the closed loop gives FID 31.42 and preference accuracy 0.693, below the full model's 29.56 and 0.734. FID over the first five closed-loop rounds is 34.2, 31.8, 30.5, 29.8 and 29.6, with per-round improvement below 1% after round five; at rule-evaluation frequency K=5 the model gives FID 29.96 and 4.3 s latency, 31.7% faster than K=1.
Perspective
All metrics are measured on a user pool of 10,000 virtual agents, outputs are 256×256 images, and both the diffusion backbone and the CART structure stay frozen; the authors have designed a pre-registered two-stage human validation, about 60 participants in-lab and about 600 online cross-culturally.
Although the synthetic data match the seed on four distributional indices, they cannot by design encode higher-order aesthetic factors such as emotional resonance, cultural contextualisation, autobiographical memory and fatigue effects; the authors note the reported gains may be over-optimistic and the baseline ranking may shift once real ratings are introduced.