Open Source Linux AI Models

VGGSfM is an advanced structure-from-motion (SfM) framework jointly developed by Meta AI Research (GenAI) and the University of Oxford’s Visual Geometry Group (VGG). It reconstructs 3D geometry, dense depth, and camera poses directly from unordered or sequential images and videos. The system combines learned feature matching and geometric optimization to generate high-quality camera calibrations, sparse/dense point clouds, and depth maps in standard COLMAP format. Version 2.0 adds support for dynamic scene handling, dense point cloud export, video-based reconstruction (1000+ frames), and integration with Gaussian Splatting pipelines. It leverages tools like PyCOLMAP, poselib, LightGlue, and PyTorch3D for feature matching, pose estimation, and visualization. With minimal configuration, users can process single scenes or full video sequences, apply motion masks to exclude moving objects, and train neural radiance or splatting models directly from reconstructed outputs.

The Video PreTraining (VPT) repository provides code and model artifacts for a project where agents learn to act by watching human gameplay videos—specifically, gameplay of Minecraft—using behavioral cloning. The idea is to learn general priors of control from large-scale, unlabeled video data, and then optionally fine-tune those priors for more goal-directed behavior via environment interaction. The repository contains demonstration models of different widths, fine-tuned variants (e.g. for building houses or early-game tasks), and inference scripts that instantiate agents from pretrained weights. Key modules include the behavioral cloning logic, the agent wrapper, and data loading pipelines (with an accessible skeleton for loading Minecraft demonstration data). The repo also includes a run_agent.py script for testing an agent interactively, and an agent.py module encapsulating the control logic.

All models (at least currently) are supported by chaiNNer, an upscaling GUI that allows for both very simple and very complex tasks to be completed in a nice manner where you "chain" nodes together. Highly recommended for images. If you're looking to upscale videos using the models then use enhancr simply due to the fact that it supports TensorRT, which will allow you to upscale videos at incredible speeds! The GUI is one of the best looking applications out there and is personally my go to option. While yes builds are paid, it is well worth your money and beats any other GUI out there currently.

Mesh R-CNN is a 3D reconstruction and object understanding framework developed by Facebook Research that extends Mask R-CNN into the 3D domain. Built on top of Detectron2 and PyTorch3D, Mesh R-CNN enables end-to-end 3D mesh prediction directly from single RGB images. The model learns to detect, segment, and reconstruct detailed 3D mesh representations of objects in natural images, bridging the gap between 2D perception and 3D understanding. Unlike voxel-based or point-based approaches, Mesh R-CNN uses a differentiable mesh representation, allowing it to efficiently refine surface geometry while maintaining high spatial detail. The system combines 2D detection from Mask R-CNN with 3D reasoning modules that output full mesh reconstructions aligned with the input image. It has been evaluated on datasets such as Pix3D, where it demonstrates state-of-the-art performance in reconstructing real-world object geometry.

TimeSformer is a vision transformer architecture for video that extends the standard attention mechanism into spatiotemporal attention. The model alternates attention along spatial and temporal dimensions (or designs variants like divided attention) so that it can capture both appearance and motion cues in video. Because the attention is global across frames, TimeSformer can reason about dependencies across long time spans, not just local neighborhoods. The official implementation in PyTorch provides configurations, pretrained models, and training scripts that make it straightforward to evaluate or fine-tune on video datasets. TimeSformer was influential in showing that pure transformer architectures—without convolutional backbones—can perform strongly on video classification tasks. Its flexible attention design allows experimenting with different factoring (spatial-then-temporal, joint, etc.) to trade off compute, memory, and accuracy.

The codebase was designed to help researchers and practitioners quickly reproduce FAIR’s results and leverage robust pre-trained backbones for downstream tasks. It also integrates Gradient Blending, an audio-visual modeling method that fuses modalities effectively (available in the Caffe2 implementation). Although VMZ is now archived and no longer actively maintained, it remains a valuable reference for understanding early large-scale video model training, transfer learning, and multimodal integration strategies that influenced modern architectures like SlowFast and X3D.