ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models

TL;DR

ViewDiff introduces a novel approach for 3D-consistent image generation using pretrained text-to-image models, integrating 3D volume rendering and cross-frame attention to produce high-quality, multi-view images from real-world data.

Contribution

The paper presents a new method that leverages pretrained text-to-image models with integrated 3D volume rendering and autoregressive generation for consistent 3D asset creation.

Findings

Produces highly 3D-consistent images from real-world data

Achieves 30% lower FID and 37% lower KID scores compared to previous methods

Generates diverse high-quality shapes and textures in authentic environments

Abstract

3D asset generation is getting massive amounts of attention, inspired by the recent success of text-guided 2D content creation. Existing text-to-3D methods use pretrained text-to-image diffusion models in an optimization problem or fine-tune them on synthetic data, which often results in non-photorealistic 3D objects without backgrounds. In this paper, we present a method that leverages pretrained text-to-image models as a prior, and learn to generate multi-view images in a single denoising process from real-world data. Concretely, we propose to integrate 3D volume-rendering and cross-frame-attention layers into each block of the existing U-Net network of the text-to-image model. Moreover, we design an autoregressive generation that renders more 3D-consistent images at any viewpoint. We train our model on real-world datasets of objects and showcase its capabilities to generate instances with a variety of high-quality shapes and textures in authentic surroundings. Compared to the existing methods, the results generated by our method are consistent, and have favorable visual quality (-30% FID, -37% KID).