The key to turning flat video into 360 panoramas is all about projection, not forcing the model to guess.

The VR-Outpaint 1.0 (an IC-LoRA for LTX2.3) is a pretty clever idea. If you just tell a video model to turn a flat shot into 360°, it has no clue how to map those rectangular pixels onto a sphere—it’s trying to solve geometry and inpainting at the same time, and it just messes up. The author’s approach is to first apply an inverse sundial projection to the input video, placing the flat footage onto a black equirectangular canvas at the correct angle and field of view.

Now the model can clearly see where the known pixels are on the sphere, and it only has to handle the inpainting part. The accompanying ComfyUI node automatically calculates the projection and uses GeoCalib to estimate the FOV, so you don’t need to measure it manually.

Sweet spot scenes are semi-static cityscapes, empty landscape shots, and slow pans. The author says the biggest lesson from this project is that a clean dataset matters more than model architecture—they rebuilt the entire dataset using CC-BY footage.

Project first, then fill in the gaps—breaking the problem down like that is a smart move.

Clean dataset matters more than architecture. Any old-school engineer knows that.

Auto FOV estimation saves you from having to measure it manually, super user-friendly.

Slow, empty shots are fine, but fast motion estimation just falls apart.

Using CC-BY to rebuild the dataset, gotta give props for that attitude.

“Reverse sundial projection” — never heard that term before, gonna keep that in mind.

Anyone got a tutorial?