Shih-Chieh Su

PASTEL (Progressive Attention-driven Stylization Through Evolving Latents) is a generative art framework that treats the internal attention pathways of pretrained latent diffusion models as a navigable style space.

Standard text-to-image models are inherently limited by language and struggle to generate complex brushwork or emergent, "unpromptable" aesthetics through textual prompts alone. To bypass these linguistic constraints, PASTEL introduces a principled "Rank-and-Update" denoising loop that dynamically evaluates information gain within the latent space. By providing various target strategies and methodologies to selectively tune targeted attention blocks, the framework transforms diffusion into a structured, traversable landscape , enabling artists to interactively steer internal layers toward high-quality aesthetic outcomes.

Latent Painting within the PASTEL framework reimagines the generative denoising loop as a dynamic canvas where the painting process actively visualizes how attention pathways are systematically emphasized over time. This approach allows users to use PASTEL's diverse block-selection strategies to drive the evolution of the latent space. As the animation progresses, it reveals the "Trails of Emphasis"—the visual footprints of how specific attention layers are being tuned and scaled along the denoising steps. By watching these hidden algorithmic mechanisms manifest as shifting brushwork, fluid textures, and evolving structural forms, the process transforms abstract model intervention into an intuitive, observable art of latent manipulation.