Persp to ERP Warp

SplatKit starts from a 360° panorama, but most of us don't have one lying around - we have a phone photo. Persp to ERP Warp is the front end that bridges the gap (it's the "Image→pano" stage in workflow 0_generate_360_panorama-upscale.json): it takes a normal perspective photo and warps it onto an equirectangular (ERP) canvas, marking everything outside the photo's footprint as an inpainting mask. You then hand control_video + inpaint_mask to an image-to-pano generator (Krea 2 Turbo, or Qwen-Image-Edit with a 360 LoRA), it invents the rest of the sphere, and you have a panorama to feed the rest of SplatKit. So: this node makes your single 360° input instead of a 360° camera.

The warp itself is area-prefiltered - the source is heavily minified into its ERP footprint, so supersample (default 2) kills aliasing/moire before a bicubic warp, and you'd go 3–4 for very high-res sources. Placement is yaw_deg/pitch_deg (0 = forward/horizon). h_fov_deg matters most: match your camera (phone main ~70, wide ~90, ultrawide ~110–120), since vertical FOV follows from the image aspect. Keep width = 2 × height - the pano LoRA expects a 2:1 canvas.

Inputs and outputs that matter

Required: image, width/height (1440×720 default), length (frames in the still batch - 1 is a single ERP still, 5 gives the video model a little temporal extent), h_fov_deg.

The optional list has two worth knowing:

• mask_feather (default 8 px) - softens the known/hole boundary so an inpainting sampler blends instead of leaving a hard seam.
• fill_color (default black) - tints the unknown region on control_video only; the masks are unchanged so inpainting still regenerates it. If a model reacts badly to black init, try #808080 mid-grey.

Outputs: control_video (the ERP canvas), control_mask (where the photo lands, for conditioning), inpaint_mask (a MASK for the unknown region - what an inpainting sampler actually uses), and erp_guide - a technical equirect reference chart (lat/lon grid, degree labels, floor/ceiling perspective curves, and a dotted FOV box showing exactly where your image sits) that you can feed a generator as a pure geometry hint for the 360 layout. Handy when the model keeps drawing walls at the wrong angle.

Install & gotchas

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, and this node needs no model downloads - the heavy MoGe/SphereSfM stuff only triggers when you reach the actual SplatKit stages. Two traps: a wrong h_fov_deg visibly warps the photo's perspective, and if your source was taken at a significant roll, straighten it before warping (workflow 0 includes a roll-180° seam fix on the output pano side for a reason). If you have no idea what your lens FOV is, run Estimate FOV first and wire its h_fov_deg in - that's exactly the pairing it was built for.