🚀 **What if 3D generation did not need a complicated cascade of models?** Most recent 3D asset generation pipelines still follow a multi-stage recipe: 🖼️ generate multi-view images → 🧱 reconstruct geometry → 🔁 convert representations → 🛠️ refine artifacts… and hope that errors do not accumulate along the way. In our new paper, **PixGS: Pixel-Space Diffusion for Direct 3D Gaussian Splat Generation**, accepted at **ECCV 2026** 🎉, we ask a simpler question: > Can we generate 3D Gaussian Splats directly — in a single stage — while preserving fine-grained geometry and appearance? Our answer is **PixGS** ✨, a pixel-space diffusion framework that directly denoises 3D Gaussian attributes instead of relying on lossy latent compression. This allows the model to regularize the generated asset at the **splat level** 🎯, improving both geometry and appearance while avoiding many decoding artifacts seen in latent/cascade pipelines. The key idea is to bring the strength of large 2D generative priors into 3D generation 🌍, but without forcing the 3D representation through a compressed latent bottleneck. PixGS uses richer supervision signals — including 🔹 RGB 🔹 depth 🔹 surface normals 🔹 high-frequency structural cues to better preserve object details and view consistency. The result: high-quality text/image-conditioned 3DGS generation with strong performance and fast inference — around **1 second on a single A100 GPU** ⚡. I’m especially excited about this direction because 3D Gaussian Splatting has already transformed how we render and represent scenes. PixGS suggests that the next bottleneck may not be rendering anymore, but how cleanly, directly, and efficiently we can **generate** 3D assets in the first place. 📄 Paper: arxiv.org/abs/2607.01803 Huge thanks to my co-author Cao Duy and everyone who supported this work! 🙌 #ECCV2026 #ComputerVision #GenerativeAI #3DGeneration #GaussianSplatting #DiffusionModels #AIResearch #Graphics #Robotics #QualcommAIResearch