barakplasma/lingbot-map-long-video

Long-video 3D reconstruction with LingBot-Map: video in, bounded GLB point cloud + camera trajectory out. Community Cog wrapper of Robbyant/lingbot-map.

Public

4 runs

License

GitHub

Playground API Examples README Versions

Run barakplasma/lingbot-map-long-video with an API

Use one of our client libraries to get started quickly. Clicking on a library will take you to the Playground tab where you can tweak different inputs, see the results, and copy the corresponding code to use in your own project.

Input schema

The fields you can use to run this model with an API. If you don't give a value for a field its default value will be used.

Field	Type	Default value	Description
video	string		The video to reconstruct into a 3D point cloud. For anything larger than a few " "seconds, pass a hosted URL rather than uploading the file directly.
sample_fps	number	5 Min: 0.25 Max: 30	How many frames per second to pull from the video. This is the main quality vs. " "speed dial: higher values give the model more overlap to work with (denser, more accurate " "geometry) but take longer and cost more. 5 suits a slow phone walkthrough; drop to 1-2 for " "fast-moving or very long footage.
max_points	integer	3000000 Min: 10000 Max: 8000000	Upper bound on the number of points in the full-resolution scene.glb. Higher means " "more detail and a larger file; lower keeps the download small. Points are spread evenly across " "the video so no section is starved.
max_preview_points	integer	50000 Min: 1000 Max: 300000	Detail of the in-browser preview (scene_preview.glb) that renders in Replicate's " "viewer. Each point becomes a small cube, so raising this sharpens the preview but enlarges the " "file and slows loading on phones. The full-resolution scene.glb is unaffected.
show_trajectory	boolean	True	Draw the camera's path through the scene as a colored tube in both GLB files.
confidence_percentile	number	60 Max: 99	Advanced. Discards this percentage of the least-confident points in every frame " "before export. Raise it to clean up noisy floaters and haze at the cost of some real geometry; " "lower it to keep more points including uncertain ones.
spatial_stride	integer	4 Min: 1 Max: 16	Advanced. Keeps every Nth pixel in each direction when turning depth maps into " "points (4 uses 1/16 of the pixels). Lower gives denser geometry and larger files; higher is " "faster and lighter. Works together with max_points, which is the final ceiling.
point_frame_stride	integer	2 Min: 1 Max: 30	Advanced. Uses every Nth sampled frame for point geometry (2 = every other frame). " "Camera poses are still recorded for every frame in trajectory.json; this only thins the points. " "Raise it to lighten very long videos.
max_frames	integer	12000 Min: 2 Max: 30000	Advanced. Hard safety cap on total frames processed, applied uniformly across the " "whole video. Acts as a backstop against runaway runtime on long footage; sample_fps usually " "reaches the limit first.
keyframe_interval	integer	2 Min: 1 Max: 30	Advanced. Persists every Nth frame in the model's KV cache. Larger values save " "memory on long, full-FPS footage (upstream suggests 8-12 there) at some accuracy cost; leave " "at 2 for ordinary clips.
window_size	integer	128 Min: 16 Max: 256	Advanced. KV-cache slots per inference window (upstream long-video recommendation " "is 128). Rarely needs changing unless you are tuning memory use.
overlap_keyframes	integer	8 Min: 2 Max: 32	Advanced. Number of keyframes shared between consecutive windows and macro-chunks " "so their poses and depths align. More overlap improves alignment continuity but adds compute.
chunk_frames	integer	512 Min: 128 Max: 1536	Advanced. Largest number of dense prediction frames held in memory at once. Lower " "it if a very long video runs out of memory; raise it for slightly better cross-chunk alignment " "when memory allows.

{
  "type": "object",
  "title": "Input",
  "required": [
    "video"
  ],
  "properties": {
    "video": {
      "type": "string",
      "title": "Video",
      "format": "uri",
      "x-order": 0,
      "description": "The video to reconstruct into a 3D point cloud. For anything larger than a few \"\n            \"seconds, pass a hosted URL rather than uploading the file directly."
    },
    "max_frames": {
      "type": "integer",
      "title": "Max Frames",
      "default": 12000,
      "maximum": 30000,
      "minimum": 2,
      "x-order": 8,
      "description": "Advanced. Hard safety cap on total frames processed, applied uniformly across the \"\n            \"whole video. Acts as a backstop against runaway runtime on long footage; sample_fps usually \"\n            \"reaches the limit first."
    },
    "max_points": {
      "type": "integer",
      "title": "Max Points",
      "default": 3000000,
      "maximum": 8000000,
      "minimum": 10000,
      "x-order": 2,
      "description": "Upper bound on the number of points in the full-resolution scene.glb. Higher means \"\n            \"more detail and a larger file; lower keeps the download small. Points are spread evenly across \"\n            \"the video so no section is starved."
    },
    "sample_fps": {
      "type": "number",
      "title": "Sample Fps",
      "default": 5,
      "maximum": 30,
      "minimum": 0.25,
      "x-order": 1,
      "description": "How many frames per second to pull from the video. This is the main quality vs. \"\n            \"speed dial: higher values give the model more overlap to work with (denser, more accurate \"\n            \"geometry) but take longer and cost more. 5 suits a slow phone walkthrough; drop to 1-2 for \"\n            \"fast-moving or very long footage."
    },
    "window_size": {
      "type": "integer",
      "title": "Window Size",
      "default": 128,
      "maximum": 256,
      "minimum": 16,
      "x-order": 10,
      "description": "Advanced. KV-cache slots per inference window (upstream long-video recommendation \"\n            \"is 128). Rarely needs changing unless you are tuning memory use."
    },
    "chunk_frames": {
      "type": "integer",
      "title": "Chunk Frames",
      "default": 512,
      "maximum": 1536,
      "minimum": 128,
      "x-order": 12,
      "description": "Advanced. Largest number of dense prediction frames held in memory at once. Lower \"\n            \"it if a very long video runs out of memory; raise it for slightly better cross-chunk alignment \"\n            \"when memory allows."
    },
    "spatial_stride": {
      "type": "integer",
      "title": "Spatial Stride",
      "default": 4,
      "maximum": 16,
      "minimum": 1,
      "x-order": 6,
      "description": "Advanced. Keeps every Nth pixel in each direction when turning depth maps into \"\n            \"points (4 uses 1/16 of the pixels). Lower gives denser geometry and larger files; higher is \"\n            \"faster and lighter. Works together with max_points, which is the final ceiling."
    },
    "show_trajectory": {
      "type": "boolean",
      "title": "Show Trajectory",
      "default": true,
      "x-order": 4,
      "description": "Draw the camera's path through the scene as a colored tube in both GLB files."
    },
    "keyframe_interval": {
      "type": "integer",
      "title": "Keyframe Interval",
      "default": 2,
      "maximum": 30,
      "minimum": 1,
      "x-order": 9,
      "description": "Advanced. Persists every Nth frame in the model's KV cache. Larger values save \"\n            \"memory on long, full-FPS footage (upstream suggests 8-12 there) at some accuracy cost; leave \"\n            \"at 2 for ordinary clips."
    },
    "overlap_keyframes": {
      "type": "integer",
      "title": "Overlap Keyframes",
      "default": 8,
      "maximum": 32,
      "minimum": 2,
      "x-order": 11,
      "description": "Advanced. Number of keyframes shared between consecutive windows and macro-chunks \"\n            \"so their poses and depths align. More overlap improves alignment continuity but adds compute."
    },
    "max_preview_points": {
      "type": "integer",
      "title": "Max Preview Points",
      "default": 50000,
      "maximum": 300000,
      "minimum": 1000,
      "x-order": 3,
      "description": "Detail of the in-browser preview (scene_preview.glb) that renders in Replicate's \"\n            \"viewer. Each point becomes a small cube, so raising this sharpens the preview but enlarges the \"\n            \"file and slows loading on phones. The full-resolution scene.glb is unaffected."
    },
    "point_frame_stride": {
      "type": "integer",
      "title": "Point Frame Stride",
      "default": 2,
      "maximum": 30,
      "minimum": 1,
      "x-order": 7,
      "description": "Advanced. Uses every Nth sampled frame for point geometry (2 = every other frame). \"\n            \"Camera poses are still recorded for every frame in trajectory.json; this only thins the points. \"\n            \"Raise it to lighten very long videos."
    },
    "confidence_percentile": {
      "type": "number",
      "title": "Confidence Percentile",
      "default": 60,
      "maximum": 99,
      "minimum": 0,
      "x-order": 5,
      "description": "Advanced. Discards this percentage of the least-confident points in every frame \"\n            \"before export. Raise it to clean up noisy floaters and haze at the cost of some real geometry; \"\n            \"lower it to keep more points including uncertain ones."
    }
  }
}

Output schema

The shape of the response you’ll get when you run this model with an API.

Schema

{
  "type": "array",
  "items": {
    "type": "string",
    "format": "uri"
  },
  "title": "Output"
}