Splatent framework enhances 3D Gaussian Splatting with diffusion latents for novel view synthesis

Researchers have introduced Splatent, a novel framework that enhances 3D Gaussian Splatting within the latent space of VAEs for improved novel view synthesis. Unlike previous methods that struggled with multi-view consistency and often resulted in blurred textures, Splatent recovers fine-grained details by processing them in 2D from input views using multi-view attention. This approach maintains the reconstruction quality of pre-trained VAEs while achieving state-of-the-art results in 3D reconstruction, offering potential for high-quality sparse-view reconstruction. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a new method for high-quality sparse-view 3D reconstruction by enhancing VAE latent spaces with multi-view attention.

RANK_REASON This is a research paper introducing a new method for 3D reconstruction.

Read on arXiv cs.CV →

• Splatent
• VAE
• 3D Gaussian Splatting
• arXiv
• Or Hirschorn

• paper
• other