FlashWorld: High-quality 3D Scene Generation within Seconds (imlixinyang.github.io)

FlashWorld is a generative model that creates high-quality 3D scenes from a single image or a text prompt in seconds by directly producing 3D Gaussian representations during multi-view generation. Unlike prior "MV-oriented" pipelines that first synthesize multiple views and then reconstruct 3D, FlashWorld embraces a 3D-oriented paradigm to guarantee geometric consistency and speed. To overcome the typical visual-quality gap of 3D-oriented methods, the authors introduce a dual-mode pre-training that learns both MV-oriented and 3D-oriented generation (bootstrapped from a video diffusion prior), followed by a cross-mode post-training distillation that aligns the 3D-oriented output distribution to the higher-fidelity MV-oriented mode. Technically, this combination preserves 3D consistency while elevating rendering quality and cutting required denoising steps at inference—enabling near real-time generation. The model also ingests large collections of single-view images and text prompts during training to improve robustness to out-of-distribution inputs. Overall, FlashWorld demonstrates a practical path to fast, consistent, and photorealistic 3D scene synthesis, with implications for interactive content creation, AR/VR asset generation, and downstream 3D modeling workflows where both speed and multi-view fidelity matter.