FlashWorld: High-quality 3D Scene Generation within Seconds

2025-10-15

AI Summary

Existing 3D scene generation methods suffer from slow inference and low rendering quality. To address these limitations, we propose the first 3D-aware diffusion-based generative framework, leveraging dual-mode (single-image or text) pretraining and cross-modal distillation to jointly optimize multiview visual fidelity and 3D geometric consistency. Our method builds upon a video diffusion prior and introduces a dual-path multiview diffusion architecture, enhanced by large-scale single-view image and text-prompt data to improve generalization. Experiments demonstrate that our approach achieves significantly improved visual quality while maintaining high 3D consistency; inference runs at second-level speed—10–100× faster than state-of-the-art methods; and supports flexible input modalities (single image or text), enabling high-fidelity, efficient 3D reconstruction across diverse scenes.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Multimodal Learning
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

We propose FlashWorld, a generative model that produces 3D scenes from a single image or text prompt in seconds, 10~100$ imes$ 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.

Problem

• Generates 3D scenes from single images or text prompts rapidly
• Improves 3D consistency while maintaining high visual quality
• Enhances generalization for out-of-distribution inputs using cross-mode training

Innovation

• Direct 3D Gaussian generation from single input
• Dual-mode pre-training with video diffusion prior
• Cross-mode distillation for quality and consistency