NVIDIA has announced 'ArtiFixer,' an AI model that generates 3D scenes from multiple photos, capable of filling in parts that are not present in the original images.
NVIDIA has developed ' ArtiFixer ,' an AI model that can generate 3D scenes from multiple photos. Based on the video generation AI
ArtiFixer: Enhancing and Extending 3D Reconstruction with Auto-Regressive Diffusion Models
https://research.nvidia.com/labs/sil/projects/artifixer/
nvidia/ArtiFixer · Hugging Face
https://huggingface.co/nvidia/ArtiFixer
3D scene reconstruction works great until the camera never sees part of the scene.
— NVIDIA AI (@NVIDIAAI) June 22, 2026
ArtiFixer from NVIDIA Research is an open autoregressive model that fills in the missing geometry that other methods leave blank. #SIGGRAPH2026 paper, code + demo: https://t.co/D9PX2OzbZf pic.twitter.com/AGQicvVKkW
Research is progressing on a technique called '3D Gaussian splatting' as a method for generating 3D scenes from multiple photographs. However, conventional 3D Gaussian splatting systems have problems such as 'often failing to maintain scene consistency' and 'the representation of parts not visible in the photograph breaking down.' ArtiFixer is an AI model with approximately 16.9 billion parameters developed based on Wan 2.1, and it has succeeded in improving the quality of 3D scenes by 'generating and inserting parts not visible in the photograph.'
ArtiFixer's training was conducted in two stages. In the first stage, it was trained to become an AI model with the function of generating and inserting parts that are not visible in the photograph. In the second stage, it underwent training that involved distilling it into an autoregressive model that generates hundreds of frames from a single frame.
There are three types of ArtiFixer:
- ' ArtiFixer ' generates views using an autoregressive model.
- ' ArtiFixer3D ' distills ArtiFixer's output into a 3D representation.
- ' ArtiFixer3D+ ' applies an autoregressive model as a post-processing step to the results of ArtiFixer3D.
The images below show 3D scenes generated using the conventional '3DGUT' method (top left), 'ArtiFixer' (top right), 'ArtiFixer3D' (bottom left), and 'ArtiFixer3D+' (bottom right). ArtiFixer can generate sharp scenes, while ArtiFixer3D is highly consistent but slightly blurry. ArtiFixer3D+ can generate both sharp and highly consistent scenes.
Below are comparison images of scenes generated using '3DGUT,' 'GenFusion,' 'GSFixer,' and 'ArtiFixer3D+.' ArtiFixer3D+ can generate 3D scenes of significantly higher quality compared to the other methods.
High-quality 3D scenes can be generated even from photographs of rooms with a large number of objects.
Related Posts:
in AI, Posted by log1o_hf