AttentionPainter: The AI That Creates Complex Art in a Single, Swift Glance

In the evolving world of computer vision, 'Stroke-based Rendering' challenges machines to decompose an image into a sequence of brushstrokes, creating a digital painting that resembles the original input. Historically, neural painting methods have relied on deep learning and reinforcement learning to predict these strokes one by one. While effective, this iterative approach often suffers from long inference times and unstable training.

Researchers have now introduced AttentionPainter, an efficient and adaptive model designed to solve these bottlenecks through single-step neural painting. Unlike previous auto-regressive methods that painstakingly predict strokes in sequence, AttentionPainter utilises a novel scalable stroke predictor. This allows the model to determine a vast number of stroke parameters within a single forward process, making it significantly faster than its predecessors.

To handle this increased speed without losing fidelity, the team developed a 'Fast Stroke Stacking' algorithm, which accelerates training by 13 times. Furthermore, they implemented a 'Stroke-density Loss' mechanism. This feature encourages the model to utilise smaller strokes when reconstructing detailed information, ensuring the final output remains sharp and accurate. Beyond mere replication, the system includes a Stroke Diffusion Model that aids in editing and inpainting, offering new tools for human artists. Extensive experiments indicate that this streamlined approach successfully outperforms state-of-the-art neural painting methods.