Create360SweepVideoFramesNode

You've got an equirectangular 360° image - maybe generated by a 360 LoRA, maybe an HDRI you downloaded - and you want a video of the camera gliding around it. The obvious move is to load it into Blender, add a camera, keyframe a rotation, and render. The obvious move is also overkill. This node is the shortcut: feed it the image, and it cranks out the individual frames of a full 360° sweep, saved straight to disk.

It ships in ProGamerGov's ComfyUI_pytorch360convert_video pack, from the same author who makes the popular Qwen 360° LoRA and the ComfyUI_preview360panorama viewer. The author himself points people here when they ask how to turn 360s into video, so this is the intended path, not a random side tool. No models to download, no API key - the whole thing is one PyTorch function.

What it actually does

Inside, the node uses the e2p function from the pytorch360convert library to re-project the equirectangular image into a perspective view, one frame at a time. start_yaw_deg to end_yaw_deg (default 0→360, i.e. a full spin) is divided into duration_per_image × fps frames, and each frame is rendered at a different yaw angle. It's a pure math operation, no diffusion involved, which is why it's fast and runs happily on CPU.

The few inputs that matter:

• images - your equirectangular image(s), straight out of whatever generated them.
• fov_h / fov_w - horizontal and vertical field of view in degrees. This is where people get burned (more below).
• width / height - output frame size.
• duration_per_image and fps - together they set your frame count. The default 5 seconds at 60fps means 300 frames per image, which is a lot of PNGs.
• total_duration - set this above 0 and it overrides duration_per_image, splitting the total evenly across all input images.
• speed_profile - constant for a uniform spin, ease_in_out if you want it to slow down at the start and end like a proper camera move.
• vertical_mode - none, during (adds a sine or linear pitch wobble while rotating), or separate (appends a dedicated up-over-the-pole sweep at the middle yaw).

Frames land in <output_dir>/video_frames/ with your filename_prefix and a zero-padded index, in whatever save_format you pick (PNG default). The node is an output node, so it doesn't return a tensor - it shows you a preview of a few frames and writes the rest.

The FOV trap

Here's the gotcha that produces subtly stretched videos: the defaults are fov_h=120, fov_w=100 with width=1920, height=1080. For undistorted perspective, your FOV ratio should match your pixel aspect ratio - fov_h / fov_w = width / height. 120/100 = 1.2, but 1920/1080 = 1.78. Result: vertical stretching. Fix it by setting fov_w so the ratio lines up - for 1080p with fov_h=120, use fov_w = 67.5. Play with the numbers; a narrower FOV gives a more telephoto, "zoomed-in" feel.

Installing it

ComfyUI Manager is the easy route - search for "pytorch360convert" (the pack title is ComfyUI_pytorch360convert_video). Or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ProGamerGov/ComfyUI_pytorch360convert_video

The only real dependency is pytorch360convert, a small PyPI package that the install pulls in automatically. Restart ComfyUI and you'll find both nodes under the pytorch360convert/video category.

Finishing the job

The node deliberately stops at frames - you assemble the video yourself with FFmpeg:

ffmpeg -framerate 60 -i video_frames/sweep360_%06d.png -c:v libx264 -pix_fmt yuv420p sweep.mp4

Common gotchas: forget the FOV ratio and your "360 tour" looks like a funhouse mirror; set 60fps/5s and suddenly you're writing 300 files per image (drop to 30fps or 2-3 seconds if you're just previewing); and remember the frames go into a video_frames subfolder, not the output root. If the node's slow, that's the CPU render path - it trades speed for not touching your VRAM, which is usually the right trade for a tool like this.