CustomNet: Zero-shot Object Customization with Variable-Viewpoints in Text-to-Image Diffusion Models

TL;DR

The paper addresses the limitations of existing object customization in diffusion models, such as time-consuming optimization and poor identity preservation, by integrating 3D novel view synthesis to enable explicit viewpoint, location, and background control. It introduces CustomNet, a zero-shot framework that leverages a viewpoint-conditioned diffusion backbone with location-aware input, dual-attention background conditioning, and a dataset construction pipeline combining synthetic multi-view data and real-world images. The approach achieves superior identity preservation and controllable viewpoint outputs without test-time optimization, demonstrated against strong baselines and via ablations. This work advances harmonized, flexible object customization in T2I diffusion models with practical implications for image compositing and personalized content creation.

Abstract

Incorporating a customized object into image generation presents an attractive feature in text-to-image generation. However, existing optimization-based and encoder-based methods are hindered by drawbacks such as time-consuming optimization, insufficient identity preservation, and a prevalent copy-pasting effect. To overcome these limitations, we introduce CustomNet, a novel object customization approach that explicitly incorporates 3D novel view synthesis capabilities into the object customization process. This integration facilitates the adjustment of spatial position relationships and viewpoints, yielding diverse outputs while effectively preserving object identity. Moreover, we introduce delicate designs to enable location control and flexible background control through textual descriptions or specific user-defined images, overcoming the limitations of existing 3D novel view synthesis methods. We further leverage a dataset construction pipeline that can better handle real-world objects and complex backgrounds. Equipped with these designs, our method facilitates zero-shot object customization without test-time optimization, offering simultaneous control over the viewpoints, location, and background. As a result, our CustomNet ensures enhanced identity preservation and generates diverse, harmonious outputs.

Figures (14)

• Figure 1: We propose CustomNet, a zero-shot customization method that can generate harmonious customized images with explicit viewpoint, location, and background controls simultaneously while ensuring identity preservation.
• Figure 2: Overview of CustomNet. CustomNet is able to simultaneously control viewpoint, location, and background in a unified framework, and can achieve harmonious customized image generation while effectively preserving object identity and texture details. The background generation can be controlled either through textual descriptions (the 'Generation' branch) or by providing a specific user-defined image (the 'Composition' branch).
• Figure 3: Qualitative comparison. Our CustomNet demonstrates superior capacities in terms of identity preservation, viewpoint control, and harmony of the customized image.
• Figure 4: Comparison to existing textual background inpainting method SD-Inpainting model and foreground object inpainting model Paint-by-Example. Our CustomNet can achieve a more harmonious output with diverse viewpoint changes while preserving identity.
• Figure 5: Explicit viewpoints control. Without the explicit viewpoint parameters $[R, T]$, a) Zero-1-to-3 tends to generate images that cannot change the viewpoint or have undesired artifacts; b) CustomNet easily obtains copy-pasting effects, even though it is trained with the multi-view dataset.