4DAnyone

ABot-Recon is a streaming 3D-reconstruction model and inference tool for reconstructing long video streams from video frames alone. At each step, it caches key-value features from the preceding 11 frames, predicts a point map in the current camera coordinate system, estimates the adjacent relative camera pose, and composes those poses sequentially into a global trajectory and point cloud without persistent learned long-range memory. It can output camera poses, relative poses, local and world point maps, colors, confidence maps, and run metadata. An optional loop-closure backend retrieves candidate revisited-frame pairs using DINOv2-SALAD descriptors, predicts relative-pose constraints with ABot-Recon, and refines the trajectory through sparse pose-graph optimization. The repository provides Python and command-line interfaces, a public model checkpoint, a demo, evaluation code, and visualization utilities. The released configuration targets Linux, Python 3.10 or later, PyTorch 2.5.1, and CUDA 12.1; source code is licensed under Apache License 2.0, while the model weights have separate terms in MODEL_LICENSE.md.

VidMap is an offline structure-from-motion system for video, developed by the Computer Vision and Geometry group (cvg). It combines temporal feature tracks, loop closures, metric depth, and global optimization to estimate camera poses, camera intrinsics, and a sparse 3D map from video or extracted frames. The Python package writes the reconstruction as a COLMAP model and supports optional browser-based HTML visualization and Rerun playback or flythrough recordings. The repository provides setup and execution workflows built around COLMAP, PyCOLMAP, PyTorch, xFormers, and GPU acceleration through CUDA and cuDSS.

3D Model Anything is an open-source iPhone LiDAR capture app, Three.js viewer, and agent skill for reconstructing rooms as editable Three.js scenes. The app records room geometry and photographs; the scan-to-threejs skill guides an AI coding agent through measurement, modeling, and matched-camera visual comparisons rather than performing a one-click conversion. The repository includes an editable office demo, browser views such as dollhouse, walking, overhead, wireframe, and knolling modes, and workflows for importing user captures. LiDAR-equipped iPhones or iPads are required for scanning, while non-LiDAR devices can import and review scans. The reusable source code and skill are MIT licensed; the example's images and textures have separate terms, and the repository does not distribute other rooms or original scan bundles.

Finger Frame AI is a browser app and Python CLI pipeline by Sophia Yang that places an AI-generated world inside a two-hand finger-frame gesture. It sends an uploaded video to Gemini Omni Flash for video-to-video restyling using a selected style or custom prompt, with alignment instructions intended to preserve framing, facial positions, expressions, blinks, gaze, and motion. MediaPipe Hand Landmarker then tracks both hands and the finger-frame quadrilateral frame by frame; the tracking pipeline uses anatomical corner ordering, spread and area gates with hysteresis, teleport rejection, velocity-adaptive smoothing, dropout holding, and presence fading. The compositor reveals the restyled video through the tracked quadrilateral and adds a dashed outline with pulsing corner markers. The browser app exports MP4 where browser recording supports it and otherwise WebM. The CLI separates restyling and compositing into `stylize.py` and `composite.py`, accepts formats readable by FFmpeg or OpenCV, and can produce H.264 MP4 while retaining the original audio track when present. Users provide their own Gemini API key; the app keeps it in the browser and sends it only to Google's API. Generation is billed per video, and the app recommends keeping uploads under about 15 MB. A no-key placeholder mode runs the tracking, compositing, and export pipeline with a hue-shifted stand-in.

JoyAI-Video-Edit is an open-source, instruction-guided system for editing live camera streams or uploaded videos as frames arrive. It processes frames causally without waiting for the complete video, requiring a predefined sequence length, or revisiting future frames, and supports subject and local edits, background replacement, style and motion changes, and reference-guided editing. Its autoregressive diffusion architecture combines an MLLM-based condition encoder, a causal video VAE, and a 16-billion-parameter multimodal diffusion transformer. The deployment uses aligned autoregressive distribution-matching distillation, long-horizon optimization, bounded KV-state inference, and deployment-oriented scheduling to reduce train–inference mismatch and temporal drift during streaming generation. The repository reports 30 FPS end-to-end throughput at 720 × 1248 in its deployment benchmark. The repository provides deployment code and model checkpoints, a local server with a browser interface, and instructions for CUDA-based inference. It is licensed under Apache 2.0.