Optimization of Automated Digital Painting Generation Model Using Convolutional Neural Networks

To address the challenges of high cost, difficulty in achieving consistent style, and insufficient real-time performance in game art and digital creation, this paper focuses on structure preservation and controllability in automated digital painting generation, proposing a controllable stylization framework led by convolutional neural networks. Methodologically, a content-style dual encoder and semantic adaptive normalization (sAdaIN / SPADE) are employed for layer-by-layer modulation. At the decoder, channel attention and skip connections are combined to stabilize boundaries and mid-level textures. Joint optimization is achieved using a composite objective consisting of perceptual / style / content consistency, edge constraints, and color distribution matching. Inference efficiency is also improved through half-precision and operator fusion. Experiments demonstrate that, under a unified evaluation protocol, the proposed method achieves consistent improvements over representative baselines in metrics such as FID, KID, LPIPS, SSIM, and Edge-SSIM, and outperforms subjective evaluations, end-to-end latency, and throughput. The controllable interface maintains stability for continuous adjustment of style intensity, brushstroke scale, and color offset. This paper contributes to the development of a convolutional generation paradigm that maintains structure, controls style, and is efficient and deployable. It also provides an evaluation and implementation path compatible with production processes, providing a replicable technical baseline for high-quality digital painting in industrial production scenarios.