Choose theme
AI Visuals
LingBot-Map
LingBot-Map reconstructs scene geometry from video, estimating camera poses and point clouds as frames arrive.
Robbyant's geometric context transformer retains references to earlier views to ground coordinates and address drift. This is reconstruction of the supplied scene, rather than generation of a new world. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
Video to scene geometry
A streaming reconstruction model with code for camera estimates, depth and point-cloud visualization.
Why it stands out
Geometry with retained context
The paper separates coordinate grounding, nearby pose references and longer trajectory memory within its attention design.
Availability
Checkpoints and two demo paths
Public weights accompany a local browser viewer and an offline rendering pipeline. Download availability and the publisher's preferred checkpoint are separate questions.
Why it matters
What makes it useful
For a recorded walkthrough, choose the output you need to inspect. The repository's demo.py opens an interactive local point-cloud viewer; its batch pipeline renders a headless flythrough MP4. The publisher uses the latter for a 25,000-frame example too long for its interactive viewer. A rendered tour and an interactive geometry view serve different viewing tasks. For a custom flythrough with several camera shots, the rendering guide warns that explicit camera-shaping flags such as --camera_mode or --smooth_window discard the YAML camera segments and rebuild the path. It recommends omitting those flags to keep the YAML sequence. Choose whether the shot sequence lives in YAML or command-line camera settings before combining a preset with overrides.
What to know
Where it fits
The September maintainer clarification says the first-version LingBot-Map-Long was phased out and recommends LingBot-Map directly after broader evaluation. The Hugging Face files still include lingbot-map-long.pt. Its presence does not make the older balanced-versus-long choice the current recommendation; the repository identifies lingbot-map.pt as its paper and demo checkpoint.
Notable points
What stands out
The publisher says default streaming does not reset state, and its inference range is limited by the distances encountered in training. Its guidance for pose collapse is to switch to windowed mode. A long input sequence and an unlimited spatial trajectory are different claims; retaining more frames alone does not remove that stated range boundary.
Before using
What to review
The repository recommends PyTorch 2.8.0 with CUDA 12.8 to match the offline renderer's Kaolin wheels. That renderer also needs rendering dependencies, ffmpeg and compiled CUDA extensions; the interactive demo has a separate setup path.
FlashInfer is recommended; the documented alternative uses PyTorch SDPA. The June 28 changelog records an SDPA cache fix, so an older checkout may differ from the current instructions.
The maintainers state that Apache-2.0 covers the code and released weights.
For recordings that show people or private spaces, confirm permission to use the footage and consider what the reconstruction or shared flythrough could reveal.
Reader fit
Who may find it relevant
It fits reconstruction researchers and technical builders working from recorded camera walkthroughs who can run the documented GPU pipeline. When reducing a long sequence, distinguish skipping input from reducing cached context. The demo's stride setting subsamples input frames; the documented keyframe interval keeps non-keyframe predictions but omits those frames from the cache. Those choices change different parts of the reconstruction workflow. Neither is simply a display setting.
Editorial note
Why LifeHubber lists it
We list LingBot-Map for reconstruction researchers who need to understand which visual history survives a long camera sequence. The paper describes keeping full image tokens for the first anchor and recent window, while compressing each older frame to six camera, anchor and register context tokens. This gives a concrete distinction between recent visual detail and the compact context retained from earlier frames, beyond simply counting input frames or choosing a keyframe interval. It does not establish constant memory use or accurate reconstruction of an arbitrary journey.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
More in AI Models
Keep browsing this category
Explore more AI model resources.
Gemma 4
google/gemma-4
A Google DeepMind Gemma 4 model family collection with public checkpoints including Gemma 4 12B, a dense multimodal model Google describes around local agentic workflows, native audio input, and encoder-free vision/audio handling.
NVIDIA Nemotron 3.5 Super VL Base
nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base-BF16
NVIDIA vision-language base checkpoint for continued training with text, document images, and video frames, with BF16 weights and a shared context ceiling up to one million tokens.
LTX-2.5
Lightricks/LTX-2.5
A Lightricks audio-video model family with gated component weights, synchronized sound, text-to-video, image-to-video, multishot generation, local Python and ComfyUI paths, fine-tuning tools, and a separate hosted API.
For project maintainers
Listed here? You can use the badge.
If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.
