ComfyUI Node
SplatPolish
Optimizes an existing world-frame splat cloud against posed video frames (L1 + D-SSIM photometric loss) using gsplat's differentiable rasterizer.
SplatPolish
- splats
- frames
- trajectory
- polished_splats
◄horizontal_fov60.00►
◄iterations300►
◄lr_xyz0.0002►
◄lr_rest0.0025►
◄lambda_l10.80►
◄lambda_dssim0.20►
◄opacity_reg0.01►
◄allow_torch_fallbackfalse►
◄deviceauto►
CategoryCamera/GSplat
Inputs (12)
| Name | Type | Default | Description |
|---|---|---|---|
| splats | GSPLAT | — | |
| frames | IMAGE | Ground-truth frames [T,H,W,3]. | |
| trajectory | TENSOR | [T,4,4] world-to-camera matrix per frame. | |
| horizontal_fov | FLOAT | 60.001–179 | — |
| iterations | INT | 3001–100000 | — |
| lr_xyz | FLOAT | 0.00020–1 | — |
| lr_rest | FLOAT | 0.00250–1 | Base learning rate for non-position parameters (3DGS-style ratios applied per group). |
| lambda_l1 | FLOAT | 0.800–10 | — |
| lambda_dssim | FLOAT | 0.200–10 | — |
| opacity_reg | FLOAT | 0.010–1 | — |
| allow_torch_fallback | BOOLEAN | false | Without gsplat+CUDA, optimize through the differentiable torch renderer at reduced resolution. EXTREMELY slow; expect minutes per 100 iterations. |
| deviceopt | COMBO | auto | 3 options: auto, cpu, cuda |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| polished_splats | GSPLAT | — |