Perspective to Pano Warp

The headline act of Mickmumpitz's panorama set, and the node the whole workflow hangs off. You give it an ordinary rectilinear photo - a phone snap, a screenshot, a render - and it projects it onto a 2:1 equirectangular canvas the way a real 360° camera would have seen it. Straight lines curve toward the edges, angles compress, and everything the camera didn't see stays a hole. That hole is the point: it's the region you're going to outpaint.

Why the warp exists

A panorama pipeline has a chicken-and-egg problem. You want to generate the 360° surroundings of a photo, but diffusion models are trained on flat rectilinear images, not equirect projections. If you just outpaint sideways from the raw photo you get a flat wide image, not a spherical one, and it won't tile when you wrap it. The warp fixes the input side: it places your photo onto the sphere correctly, and hands you the masks a sampler needs to fill the rest. The output even tiles left-to-right, because that's what "equirectangular" means.

How it works

It's a pinhole-to-ERP forward warp done properly in PyTorch. For every pixel of the ERP canvas the node computes a viewing ray (longitude/latitude), rotates that ray by your yaw_deg / pitch_deg, then projects it back into the source image using a focal length derived from h_fov_deg. The source gets area-downsampled to its on-canvas footprint first (supersample, default 2) so the warp resamples near 1:1 - that's what kills the moiré you'd otherwise get. Known pixels keep the photo; unknown pixels get fill_color (black by default).

The inputs that matter

• h_fov_deg - the horizontal field of view of your source photo. Photos don't store this, so you must supply it. Phone main ~70, wide ~90, ultrawide ~110–120. Too small and the photo sits in a tight window; too big and it smears across half the canvas.
• width / height - keep width = 2×height for a full pano (defaults 1408×704 are a good start).
• length - frames in the still batch (N = 1+4k). 1 for a single still; 5 gives a video model a little temporal extent.
• pitch_deg - camera tilt. If the horizon comes out wrong, this is the first value to change.
• fill_color - only tints the control_video; the masks are untouched, so an inpainting sampler still regenerates the region. If your model reacts badly to a black init, try #808080.

The outputs and where they go

• control_video - your photo placed on the sphere (the "plate"). Feeds the generator as a reference/control input.
• inpaint_mask - white = hole to generate, black = your photo. This goes straight into an inpainting sampler.
• control_mask - the same info as an RGB image, if your model wants a mask as an image.
• erp_guide - a technical chart, no photo: lat/lon grid, degree labels, floor/ceiling perspective curves, and a dotted box showing exactly where your photo sits. Feed it to the generator as a pure geometry hint for the 360° layout.

One subtlety worth knowing: mask_feather ramps into the known side only, keeping every true hole fully white. The comment in the source says a ramp there previously let the sampler blend the black init latent back in and left a dark ring - that's the bug this avoids.

Install and gotchas

It's part of Mickmumpitz-Nodes (MIT), and needs no model downloads - the pack's dependencies are just numpy, Pillow and opencv-python. Install via ComfyUI Manager (search "Mickmumpitz") or:

cd ComfyUI/custom_nodes
git clone https://github.com/mickmumpitz/ComfyUI-Mickmumpitz-Nodes.git

Restart ComfyUI and the node lives under Mickmumpitz/Panorama as "Perspective to Pano Warp".

The usual failure is a wrong FOV, which you'll see as the photo either huddled in the middle or stretched past the edges - fix it by typing your camera's real FOV rather than trusting a guess. If the image is tilted or the horizon's off, that's pitch_deg. Everything else is the fun part: feed control_video and inpaint_mask to an outpainting pass, then clean up with the pack's Harmonize Boundary and Seam Roll nodes.