The research team led by Prof. Harry YANG, Assistant Professor in the Division of Arts and Machine Creativity (AMC), has received the “CVPR Compute Transparency Champion” award at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 for their paper “DenDiff: Density-Guided Diffusion for Quantity-Aware Image Synthesis”—the highest recognition in CVPR 2026's Compute Reporting Initiative.

DenDiff addresses a common limitation of text-to-image diffusion models, which often struggle to generate the requested number of objects because numerical concepts are discrete and require explicit spatial grounding. To overcome this challenge, the method reformulates count control as a continuous density-guided generation problem. Its perception module leverages fine-grained visual representations, including DINO features, to generate spatially coherent density maps, which are then used during the denoising process to iteratively guide image generation toward the target quantity distribution while steering it away from incorrect counts. Experiments on the FSC-147 and ShanghaiTech datasets demonstrate state-of-the-art numerical accuracy while maintaining strong visual fidelity. By bridging semantic image generation with explicit numerical and spatial control, DenDiff enables more reliable quantity-aware image synthesis.

CVPR is widely regarded as one of the world’s premier conferences in computer vision and artificial intelligence. This Award specifically recognizes the paper for its exemplary reporting of computational resources, methodologies, and reproducibility details.

Congratulations to Prof. Yang and the team on receiving this international recognition!