Dense Prediction with Image Generation Models

Zanyi Wang · hf · 2026-07-13

This paper proposes a novel approach adapting pre-trained text-to-image diffusion Transformers for dense prediction.

The core idea is to bypass RGB image generation entirely. Instead, tokens are mapped directly to the task's native outputs, enabling pixel-level dense regression or classification. The authors note that this method requires minimal additional parameters yet achieves state-of-the-art performance across multiple dense prediction tasks.

Original post →

More from Research

Research channel →