Plot Camera

This is the flagship node of SplatKit - the one that turns "here's a panorama" into "here's a camera path moving through a 3D scene." One 360° pano is a single viewpoint, and you can't reconstruct 3D from one viewpoint. So Plot Camera invents the missing viewpoints: MoGe estimates depth, the pano becomes a mesh, you draw a path on the panorama, and the node renders an equirectangular control video plus a validity mask along that path. The control video is what you feed to WAN (with the Matrix-3D pano LoRA) so the video model fills the disocclusion holes with temporally coherent content. Everything after - SphereSfM, COLMAP, the trained splat - descends from this one render.

The best part is the editor. Open the graph and the panorama shows a MoGe point cloud overlay; you drag anchor points for the camera to fly through - the author calls it WYSIWYG, and it is. Anchors are literal coordinates in the same scene units as the overlay (+Z forward into the pano, +X right, +Y up, origin = start camera), one x, y, z per line or as JSON. There's deliberately no collision guard: the camera goes exactly where you put it, so yes, you can fly into a wall. Keep moves small - a single pano only holds so much parallax, and big sweeps mean big holes for WAN to invent.

The inputs that matter

• panorama - your equirect image. Feed the same image you'll feed WAN, so everything lines up.
• anchors - at least two points. This is where you spend your time.
• orientation - look_forward (faces the path tangent, default), look_at_target (every frame aims at one shared point you set with the orange "look" marker), or per_point_look (each anchor has its own target, the aim sweeps between them).
• length - frame count; must match your Wan Conditioning length (default 81). Nudge it and nothing downstream matches.
• moge_level - geometry detail from the depth network, 0–9. 6 is a balanced default; drop to 3–4 while placing anchors for speed, raise to 9 for a final sharp mesh.

Outputs: control_video and control_mask (the WAN conditioning pair), condition_dir (cameras.npz + depth/mask sidecars, a drop-in for Wan Conditioning + SphereSfM), a matplotlib camera_preview (top-down X-Z and side Z-Y plots of the path - sanity-check before committing to a WAN pass), and rail_json - the exact camera path. Wire rail_json into HiRes Composite, not condition_dir. Several plot nodes sharing one dataset_dir all write condition/, so only the last one's rail survives there; rail_json is this node's own per-node copy.

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

Restart. First run auto-downloads the MoGe checkpoint (~1.2 GB into ComfyUI/models/MoGe) and the SphereSfM binary is fetched by the SfM nodes, not this one. You supply the WAN 2.1 i2v checkpoint and the pano LoRA (converted with tools/convert_pano_lora.py → pano_video_gen_720p_comfy.safetensors in your loras folder).

Gotchas

• The prompt matters. It must describe the actual scene in the panorama; a wrong prompt visibly degrades what WAN paints into the holes. This is the #1 quality lever in the whole pipeline.
• Static paths won't triangulate. SphereSfM is real SfM and needs parallax - orbit, push-in, or spiral trajectories work; a pure pan does nothing.
• If the JS editor fails to load, the node still renders - the path is just a text widget. And note that moge_level only sets how much detail the depth network sees per sub-view; it never affects output resolution, so drop it to 3–4 while placing anchors and raise it back for the final render.