Unleashing Text-to-Image Diffusion Models for Visual Perception

TL;DR

The paper addresses the challenge of leveraging large, vision-language pre-trained text-to-image diffusion models for downstream visual perception tasks. It introduces Visual Perception with a pre-trained Diffusion model (VPD), which uses the diffusion model as a semantic backbone, injects task labels through prompting and a text adapter, and utilizes cross-attention maps as explicit guidance. Across semantic segmentation, referring image segmentation, and depth estimation, VPD achieves competitive or state-of-the-art results and demonstrates faster adaptation with shorter training schedules. The work suggests that generative pre-training can serve as a powerful, scalable source of visual representations for perception tasks, bridging image generation and understanding in a unified framework.

Abstract

Diffusion models (DMs) have become the new trend of generative models and have demonstrated a powerful ability of conditional synthesis. Among those, text-to-image diffusion models pre-trained on large-scale image-text pairs are highly controllable by customizable prompts. Unlike the unconditional generative models that focus on low-level attributes and details, text-to-image diffusion models contain more high-level knowledge thanks to the vision-language pre-training. In this paper, we propose VPD (Visual Perception with a pre-trained Diffusion model), a new framework that exploits the semantic information of a pre-trained text-to-image diffusion model in visual perception tasks. Instead of using the pre-trained denoising autoencoder in a diffusion-based pipeline, we simply use it as a backbone and aim to study how to take full advantage of the learned knowledge. Specifically, we prompt the denoising decoder with proper textual inputs and refine the text features with an adapter, leading to a better alignment to the pre-trained stage and making the visual contents interact with the text prompts. We also propose to utilize the cross-attention maps between the visual features and the text features to provide explicit guidance. Compared with other pre-training methods, we show that vision-language pre-trained diffusion models can be faster adapted to downstream visual perception tasks using the proposed VPD. Extensive experiments on semantic segmentation, referring image segmentation and depth estimation demonstrates the effectiveness of our method. Notably, VPD attains 0.254 RMSE on NYUv2 depth estimation and 73.3% oIoU on RefCOCO-val referring image segmentation, establishing new records on these two benchmarks. Code is available at https://github.com/wl-zhao/VPD

Figures (3)

• Figure 1: The main idea of the proposed VPD framework. Motivated by the compelling generative semantic of a text-to-image diffusion model, we proposed a new framework named VPD to exploit the pre-trained knowledge in the denoising UNet to provide semantic guidance for downstream visual perception tasks.
• Figure 2: The overall framework of VPD. To better exploit the semantic knowledge learned from text-to-image generation pre-training, we prompt the denoising UNet with properly designed text prompts and employ the cross-attention maps to provide both implicit and explicit guidance to downstream visual perception tasks. Our framework can fully leverage both the low-level and high-level pre-trained knowledge and can be applied in a variety of visual perception tasks.
• Figure 3: Longer pre-training yields better performance on downstream tasks. We train VPD with different versions of Stable-Diffusion (indicated by SD-1-x) on ADE20K and investigate how the pre-training iteration would affect the performance. The upward trend demonstrates that our VPD can benefit from a stronger text-to-image diffusion model.