Target Attribute Diffusion Models

William Loh, Yanting Miao, Pascal Poupart, Suraj Kothawade; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 6952-6960

Diffusion models have shown notable success in generating images conditioned on textual prompts, enabling users to edit images at a coarse scale with well-aligned text-to-image models. ControlNet enhances these capabilities by allowing diffusion models to edit aspects such as pose, position, and edges according to reference visual motion information in a qualitative manner. However, diffusion models still face challenges in measurable and quantitative applications, such as applying sharpening or color enhancement effects. We call quantities such as brightness and saturation, attributes. In this work, we introduce Target-Attribute Diffusion Models (TADM), which enable diffusion models to incorporate additional conditioning on continuous random variables. Unlike classifier-guidance methods, which require training an explicit classifier, TADM supports real-valued conditional variables. We also propose a new architecture called attribute carrier between the text embeddings and the new conditioning variable. Experiments were conducted on three attributes: color saturation, sharpness and human preference. TADM outperformed the baseline algorithm on a single prompt, single attribute experiment. In addition, TADM demonstrates improvement in the multiple prompt experiments with respect to two of the three attributes.