facebookresearch/pixio

[CVPR 2026] Pixio: a capable vision encoder dedicated to dense prediction, simply by pixel reconstruction

What it solves

Pixio provides a vision encoder specifically optimized for dense prediction tasks—such as monocular depth estimation, semantic segmentation, and 3D reconstruction—where high-resolution, pixel-level detail is critical. It addresses the limitations of general-purpose vision encoders by focusing on pixel reconstruction during pre-training.

How it works

Built upon the Masked Autoencoder (MAE) architecture, Pixio introduces three primary algorithmic improvements: a deeper decoder, larger masking granularity, and an increased number of class tokens. Additionally, it replaces the standard ImageNet-1K pre-training data with a self-curated version of the MetaCLIP-2B dataset to improve scale and quality.

Who it’s for

It is designed for researchers and developers working on computer vision tasks that require dense, pixel-accurate predictions rather than simple image-level classification.

Highlights

  • Superior Dense Prediction: Outperforms MAE, DINOv2, and DINOv3 in monocular depth estimation and 3D reconstruction.
  • Strong Segmentation Performance: Achieves competitive or leading results on benchmarks like ADE20K and Pascal VOC.
  • Scalable Architecture: Offers pre-trained models ranging from 86M (Pixio-B) to 5.4B (Pixio-5B) parameters.
  • Flexible Integration: Compatible with both raw source code and the Hugging Face Transformers API.

Related

  • Project
    EvolvingLMMs-Lab/NEO

    NEO is a series of native vision-language models using an encoder-free dense architecture to unify pixel and word processing for efficient multimodal understanding.

  • Project
    UCSC-VLAA/OpenVision

    A family of open-source, cost-effective vision encoders for multimodal learning that optimizes training efficiency and performance across understanding and generation tasks.

  • Project
    jiafeng5513/Evision

    A binocular vision system for camera calibration, disparity mapping, and 3D reconstruction using algorithms like ELAS and ADCensus.

  • Project
    nv-tlabs/PiD

    PiD is a plug-and-play diffusion decoder that replaces standard VAE/RAE decoders to turn latent representations directly into super-resolved 2K to 4K pixels in a single pass.

  • Project
    RenderKit/oidn

    A high-performance library of deep learning-based denoising filters that reduces rendering times for ray tracing applications by removing Monte Carlo noise.