Text2NeRF: Text-Driven 3D Scene Generation with Neural Radiance Fields

TL;DR

Text2NeRF tackles zero-shot text-driven 3D scene generation by integrating Neural Radiance Fields with pre-trained text-to-image diffusion models and monocular depth priors. It introduces a progressive inpainting and updating (PIU) strategy, a two-stage depth alignment, and a support-set mechanism to maintain multi-view consistency while expanding scenes from a text prompt. The approach yields photorealistic, multi-view coherent outdoor and indoor scenes and outperforms prior CLIP/diffusion-guided methods in both quality metrics and visual fidelity. While offering strong capabilities, it notes limitations in handling large occlusions and longer optimization times, pointing to future improvements in depth estimation and efficiency.

Abstract

Text-driven 3D scene generation is widely applicable to video gaming, film industry, and metaverse applications that have a large demand for 3D scenes. However, existing text-to-3D generation methods are limited to producing 3D objects with simple geometries and dreamlike styles that lack realism. In this work, we present Text2NeRF, which is able to generate a wide range of 3D scenes with complicated geometric structures and high-fidelity textures purely from a text prompt. To this end, we adopt NeRF as the 3D representation and leverage a pre-trained text-to-image diffusion model to constrain the 3D reconstruction of the NeRF to reflect the scene description. Specifically, we employ the diffusion model to infer the text-related image as the content prior and use a monocular depth estimation method to offer the geometric prior. Both content and geometric priors are utilized to update the NeRF model. To guarantee textured and geometric consistency between different views, we introduce a progressive scene inpainting and updating strategy for novel view synthesis of the scene. Our method requires no additional training data but only a natural language description of the scene as the input. Extensive experiments demonstrate that our Text2NeRF outperforms existing methods in producing photo-realistic, multi-view consistent, and diverse 3D scenes from a variety of natural language prompts. Our code is available at https://github.com/eckertzhang/Text2NeRF.

Figures (14)

• Figure 1: We propose Text2NeRF, a text-driven 3D scene generation framework by combining the neural radiance field representation and a pre-trained text-to-image diffusion model. Our Text2NeRF is capable of generating diverse and view-consistent indoor/outdoor 3D scenes solely from natural language descriptions. Please refer to our supplementary video demo for more examples.
• Figure 2: Overview of our Text2NeRF. Given an input text prompt, we infer an initial view $I_0$ and estimate its depth $D_0$ via a pre-trained diffusion model and a depth estimation model. Then we use the depth image-based rendering (DIBR) to warp the initial view and its depth map to various views to build the support set $\mathbf{S}_0$ for initializing the neural radiance field (NeRF). Afterward, we design a progressive inpainting and updating (PIU) strategy to complete missing regions consistently. During each update, we render the NeRF in a novel view $k$ to produce the image $I_k^R$ and depth $D_k^R$ with missing regions. Then, the diffusion model and depth estimation model are adopted to deduce completed image $\hat{I}_k$ and its depth $D_k^E$. Furthermore, a two-stage depth alignment is implemented on $D_k^R$ and $D_k^E$ to obtain aligned depth $\hat{D}_k$. Finally, the support set $\mathbf{S}_k$ of view $k$ is added into training data to update NeRF.
• Figure 3: Example of scale and value differences. $x_1$ and $x_2$ are two aligned pixels in different views. The spacial points $x_1^E$ and $x_2^E$ are projected based on the estimated depth $D_k^E$ in view $k$. $x_1^R$ and $x_2^R$ are points projected according to the rendered depth $D_k^R$ which is constrained by known views such as view 0. Here, $\Delta_E \neq \Delta_R$ indicates the scale difference, and $\Delta_1 \neq 0$ or $\Delta_2 \neq 0$ indicate the value difference.
• Figure 4: Examples of two inpainting strategies. The intuitive independent inpainting strategy simultaneously performs rendering and inpainting for each view. Due to there is no 3D constraint during 2D inpainting, the overlapping regions inpainted in different views will be view-inconsistent, as shown in the red box. In contrast, our progressive inpainting strategy achieves view-consistent inpainting results by introducing NeRF as a 3D constraint and reflecting previously inpainted content into the next view.
• Figure 5: Qualitative comparison of results generated by baselines and ours on different text prompts. Here, we only show two rendering results from different views for each generated scene of each method due to space limitations. Please refer to the supplementary material for video results.