FlashWorld: High-quality 3D Scene Generation within Seconds

TL;DR

FlashWorld tackles fast, high-quality 3D scene generation by integrating a dual-mode diffusion framework that supports both MV-oriented and 3D-oriented generation. A cross-mode post-training stage distills high-fidelity MV outputs into a 3D-consistent, few-step generator, aided by out-of-distribution co-training to boost generalization. The approach achieves state-of-the-art results on image-to-3D and text-to-3D tasks and excels in the WorldScore benchmark, delivering superior quality with significantly reduced inference time. Overall, the method provides a unified, efficient pipeline for versatile 3D scene synthesis from diverse inputs.

Abstract

We propose FlashWorld, a generative model that produces 3D scenes from a single image or text prompt in seconds, 10~100$\times$ faster than previous works while possessing superior rendering quality. Our approach shifts from the conventional multi-view-oriented (MV-oriented) paradigm, which generates multi-view images for subsequent 3D reconstruction, to a 3D-oriented approach where the model directly produces 3D Gaussian representations during multi-view generation. While ensuring 3D consistency, 3D-oriented method typically suffers poor visual quality. FlashWorld includes a dual-mode pre-training phase followed by a cross-mode post-training phase, effectively integrating the strengths of both paradigms. Specifically, leveraging the prior from a video diffusion model, we first pre-train a dual-mode multi-view diffusion model, which jointly supports MV-oriented and 3D-oriented generation modes. To bridge the quality gap in 3D-oriented generation, we further propose a cross-mode post-training distillation by matching distribution from consistent 3D-oriented mode to high-quality MV-oriented mode. This not only enhances visual quality while maintaining 3D consistency, but also reduces the required denoising steps for inference. Also, we propose a strategy to leverage massive single-view images and text prompts during this process to enhance the model's generalization to out-of-distribution inputs. Extensive experiments demonstrate the superiority and efficiency of our method.

Figures (9)

• Figure 1: FlashWorld enables fast and high-quality 3D scene generation across diverse scenes.
• Figure 2: A brief comparison of different 3D scene generation methods. MV-oriented diffusion methods (i.e., CAT3D gao2024cat3d, Bolt3D szymanowicz2025bolt3d, Wonderland liang2025wonderland, and Ours w/ MV-Diff) suffer from noisy textures due to multi-view inconsistency. MV-oriented distillation further exacerbates this flaw (i.e., Ours w/ MV-Dist). 3D-oriented diffusion methods (i.e., Ours w/ 3D-Diff) suffer from blurry visual effect. Our cross-mode distillation model (i.e., Ours) simultaneously solves these, making the quality of the novel view close to the input view. The time cost per scene, tested on a single GPU, is presented at the bottom of each method.
• Figure 3: Method overview. We first pre-train a dual-mode multi-view latent diffusion model using multi-view datasets, and then employ an cross-mode distillation post-training strategy to accelerate generation while enhancing visual quality and inheriting 3D consistency.
• Figure 4: Image-to-3D scene generation results of different methods.
• Figure 5: Text-to-3D scene generation results of different methods.