ComfyUI Node
VideoToFusedSplats
Runs SHARP on video keyframes, optionally scale-aligns to metric depth and filters dynamic pixels, transforms each keyframe splat cloud to the world frame via the inverse camera pose, and fuses everything incrementally into a single world-frame splat cloud.
VideoToFusedSplats
- frames
- trajectory
- static_mask
- depths
- splats
◄horizontal_fov60.00►
◄checkpoint<download default>►
◄keyframe_stride8►
◄stitch_voxel_size0.010►
◄stitch_modesmart►
◄deviceauto►
CategoryCamera/GSplat
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| frames | IMAGE | Video frames [T,H,W,3]. | |
| trajectory | TENSOR | [T,4,4] world-to-camera matrix per frame (a single [4,4] is broadcast). | |
| horizontal_fov | FLOAT | 60.001–179 | — |
| checkpoint | COMBO | <download default> | SHARP .pt checkpoint from the input folder, or download the default model. |
| keyframe_stride | INT | 81–1000 | Run SHARP on every Nth frame. |
| stitch_voxel_size | FLOAT | 0.0100–10 | — |
| stitch_mode | COMBO | smart | 4 options: smart, average, discard, keep |
| static_maskopt | MASK | [T,H,W], 1 = static/keep pixel. Splats whose source pixel has mask < 0.5 are dropped. | |
| depthsopt | TENSOR | [T,H,W] metric depths. SHARP splats are scale-aligned per keyframe via a robust median disparity ratio. | |
| deviceopt | COMBO | auto | 3 options: auto, cpu, cuda |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| splats | GSPLAT | — |