ComfyUI-SplatKit

ComfyUI-SplatKit

Build 3D Gaussian Splat training datasets from a single 360° panorama — entirely inside ComfyUI.

Feed it one equirectangular panorama and a prompt. You get back a COLMAP dataset (images/ + sparse/0 + an init point cloud) that trains straight away in LichtFeld Studio, gsplat, 3DGRUT or anything else that reads COLMAP. No external venv, nothing to build.

SplatKit produces datasets, not trained splats — training stays in whichever trainer you already like.

How it works

One panorama is a single viewpoint, and one viewpoint cannot constrain a 3D scene. So the pack invents the missing viewpoints, then reconstructs a real camera solution from them:

panorama ─▶ MoGe depth ─▶ camera-motion control video ─▶ WAN fills the disocclusions
         ─▶ SphereSfM (classical SfM) ─▶ COLMAP dataset ─▶ your trainer

1. MoGe estimates depth for the pano and turns it into a mesh.
2. A camera path you draw in the graph is rendered through that mesh → an equirect control video plus a validity mask (the holes are the parts the pano never saw).
3. WAN (i2v, with the Matrix-3D pano LoRA) fills those holes with temporally coherent content → a real moving-camera 360° video.
4. SphereSfM runs classical structure-from-motion on those panoramas and writes a COLMAP reconstruction — real matches, poses and sparse cloud.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/mickmumpitz/ComfyUI-SplatKit
python_embeded\python.exe -m pip install -r ComfyUI-SplatKit/requirements.txt

(Non-portable install: python -m pip install -r ComfyUI-SplatKit/requirements.txt.) Restart ComfyUI. Three things download on first use:

• the MoGe checkpoint → ComfyUI/models/MoGe
• the SphereSfM binary (colmap_sphere.exe, ~37 MB, SHA-256 verified) → bin/. A CUDA static-linked build of SphereSfM, so there's no CUDA toolkit to install — see docs/SPHERESFM.md. Windows + NVIDIA (Turing or newer) only; on Linux/macOS the SphereSfM nodes won't run, the rest of the pack is platform-independent.
• the RAFT optical-flow weights, the first time HiRes Composite runs with base_mode=wan.

You supply: a WAN 2.1 i2v checkpoint, and the Matrix-3D pano LoRA converted to ComfyUI's key convention with tools/convert_pano_lora.py (→ pano_video_gen_720p_comfy.safetensors in your loras folder).

Workflows

Ready-made graphs in workflows/. Start with 1 — or 0 if you don't have a panorama yet.

| Graph | What it does | Also needs | |---|---|---| | 0_generate_360_panorama-upscale.json | Make the input pano. text→pano (Krea 2 Turbo) or image→pano (Qwen-Image-Edit + a 360 LoRA), both with detail-refine, a roll-180° seam fix, and an upscale tail. | comfyui-LatLong, ComfyUI_essentials, ComfyUI_UltimateSDUpscale, ComfyUI-Mickmumpitz-Nodes | | 1_generate-dataset-hires.json | The main graph: pano → trained-splat-ready dataset in one queue. Draw a camera path, WAN fills the fly-through, HiRes Composite reprojects the original pano through the same geometry (+115% detail indoors / +338% outdoors — docs/HIRES_COMPOSITE.md), and dual-res SphereSfM writes the COLMAP dataset. | — |

Everything else these graphs use is core ComfyUI. Install the right-hand packs only for the workflows you run — the node pack itself depends on none of them.

The prompt matters. It must describe the actual scene in the panorama — a wrong prompt visibly degrades what WAN paints into the holes.

Nodes

Nineteen nodes, all under the SplatKit category; every registered class is used.

• Core — Dataset Project, MoGe Model Loader, Camera Plot Fly-Through (Geometry), Camera Plot Scene Reference, Wan I2V Masked-Video Conditioning.
• Dataset builders — SphereSfM Dataset (recommended: classical SfM → COLMAP), SphereSfM Dataset (Dual-Res), SphereSfM Add Camera Path.
• Hi-res — HiRes Pano Fly-Through, Add HiRes Views to Dataset, HiRes Composite.
• Upscaling — Resolve Dataset Images, Load Dataset Images (Ordered), Save Upscaled Dataset, Save Upscaled Frames (Streaming).
• Image→pano — Persp to ERP Warp, Estimate FOV, Switch (workflow 0).
• Repair — Rebuild COLMAP Sparse reassembles sparse/0 from _spheresfm_work/ without re-running SfM.

The interactive path editor (web/camera_plot_geo.js) lets you drag anchors on the panorama and renders the MoGe cloud behind the path, so you can see if you're about to fly through a wall. If the JS fails to load the node still works — the path is just a text widget.

Getting sharp splats

A single panorama caps splat sharpness two ways; SplatKit gives you a lever for each. Both are optional refinements on top of the base pipeline — full detail in docs/HIRES_COMPOSITE.md.

• HiRes Pano Fly-Through renders pinhole views directly from the MoGe mesh at any resolution, taking colour from the untouched full-res panorama — so geometry and texture resolution are independent (an 8K pano lands every pixel in a 4K render). It closes disocclusions instead of punching them out for WAN, and emits a splat_mask (white = real detail, black = synthesized). Add HiRes Views to Dataset registers those renders into an existing dataset as their own PINHOLE cameras, pinning existing poses so the add can't disturb a dataset that already trains.
• HiRes Composite keeps the WAN clip but stops it repainting pixels that were never in question: it reprojects the original 8192×4096 pano through the same geometry and only lets WAN fill where geometry has no answer. Measured vs the base pipeline: eval PSNR +1.31 dB indoors / +2.37 dB outdoors, reconstructed detail 31.9%→57.5% and 25.4%→61.4%.

Training the dataset

The COLMAP output is ordinary — point any trainer at it:

LichtFeld-Studio.exe -d output/<name> -o <out> --headless --train \
  --strategy mcmc --max-cap 2000000 --sh-degree 2

An equirect dataset (Build Equirect Dataset) needs --gut. For HiRes Composite datasets, train with --max-cap 3000000 — LichtFeld's 1M default hides the resolution gain.

Rasterizer: Triton / pure-torch, no nvdiffrast

Matrix-3D's renderer imports nvdiffrast.torch. SplatKit ships its own API-compatible rasterizer instead (shim/), so there's nothing to compile and no NVIDIA-licensed dependency. Two backends, auto-selected per machine (override with P2S_RASTER_BACKEND=torch|triton):

1. Triton (shim/raster_triton.py) — in-repo GPU fast path, JIT-compiled at runtime against your own torch/CUDA (Linux torch bundles triton; on Windows: pip install triton-windows). Self-tested against the torch oracle on first use; any failure silently falls back.
2. pure torch (shim/raster_torch.py) — zero dependencies, runs everywhere (CPU / AMD / Mac).

The shim never delegates to a real nvdiffrast build, even if one is importable. Validated vs a trimesh ray-cast oracle (coverage 100%, colour MAE 0.0); 49-frame fly-through at 2048×1024 on a 5090: 4.9 s torch / 2.6 s triton.

Repo layout

__init__.py            re-exports the mappings from nodes/
prestartup_script.py   OpenEXR codec enable, run pre-import by ComfyUI
nodes/                 the ComfyUI layer — INPUT_TYPES, tensor unpacking, thin calls
core/                  the engine, no ComfyUI imports — SfM, MoGe/mesh render, reprojection
shim/                  pure-torch / triton nvdiffrast replacement
vendored/              third-party source: MoGe, utils3d, Matrix-3D utils
web/                   in-graph camera path editor (JS)
tools/                 standalone maintenance scripts
tests/                 rasterizer + planner checks, no ComfyUI needed
workflows/             the graphs above

Only __init__.py and prestartup_script.py sit at the root — ComfyUI hard-codes both locations. Module names inside core/ are deliberately distinctive (gpu_lsmr, not solve.py) because the vendored tree reaches them by bare name off the sys.path entry matrix3d_pipeline.setup_paths() adds.

License

MIT — see LICENSE. Bundled third-party code keeps its own license and notice files alongside it (vendored/, docs/).