thesis
What Words Cannot Say. Learning Precise Concepts for Image Generation.
PhD thesis · Universitat Autònoma de Barcelona · in preparation
Text-to-image diffusion models are conditioned on language, so their controllability is bounded by what language can express. Color names cover broad ranges of shades, hex codes and RGB values are broken by the tokenizer into pieces without color meaning, and continuous attributes such as age are described by only a handful of discrete words. My thesis studies how to learn, evaluate, and control such precise visual concepts.
Chapters
Chapter 1 introduces the problem and Chapter 6 concludes. The four core chapters build on each other.
2 Learning colors as prompts ColorPeel learns a dedicated token for a user-chosen color by disentangling color from shape. ECCV 2024. 3 Measuring color understanding GenColorBench evaluates color generation with color naming systems and perceptual metrics instead of VLMs. CVPR 2026. 4 Controlling color by number NumColor maps hex and RGB codes to embeddings aligned with CIE Lab, transferring zero-shot across models. ECCV 2026.
5 Steering continuous concepts Beyond color, we study how continuous visual attributes are represented inside diffusion transformers and how to steer along them. Under review.