3D Scene

ComfyTV's 3D Scene stage is the one where you stop and go "okay, this pack is serious." It's a miniature DCC - characters, primitives, lights, and a camera - running as a node body with a live 3D viewport, under the ComfyTV/Input category. The point isn't to replace Blender; it's to put a 3D stage inside your ComfyUI flow so a scene can be captured into images and video that the rest of the graph then consumes. Want a camera orbit shot of a character rigged up from generated assets, feeding straight into your video pipeline? That's the workflow this node exists for.

What it does

You build the scene in the editor (scene_state - the JSON for characters/primitives/lights/camera/output, hidden plumbing set by the viewport). The stage carries its own multi-camera setup with keyframed camera paths, and the width/height inputs (64–4096, default 1024²) set the capture resolution and lock the editor viewport's aspect, so what you frame is what you get. The channel combo picks which render channel a capture/record writes - color, depth, normal, openpose, or id - which is the killer feature: you can capture a depth pass or an OpenPose skeleton of your own 3D scene and feed it into ControlNet-style workflows without hunting for a separate tool. Each of those channel/capture choices is "internal - driven by the node body," which is ComfyTV-speak for "set it in the UI, don't type it."

The outputs are what make it a stage instead of a toy: image (the last single Capture), images (a batch, one per scene camera), and video (the last Record). All three are snapshots the downstream graph consumes without re-rendering the scene.

Should you use it?

Be honest about the trade. This is a real 3D editor - there's a learning curve to placing lights and keying a camera path, and it's not going to compete with dedicated DCCs for serious modeling. But for the ComfyUI-specific job it's aimed at - generate a 3D model, pose it, capture multi-channel renders, feed them to image/video stages - it collapses what would otherwise be five disconnected tools into one node. The depth/openpose/id channels are the reason to reach for it over just screenshotting a viewport.

Install

With the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/jtydhr88/ComfyTV

Restart ComfyUI; find it under ComfyTV → Input. ComfyUI Manager users search "ComfyTV". Desktop/macOS installs: clone into the running instance's custom_nodes (get its absolute path from the startup log; watch for a nested ComfyTV/ComfyTV/ folder) and fully restart the backend. Rendering here is CPU/GPU local math - the pack ships zero Python dependencies - but the 3D content (models, textures) is on you: loaders come from elsewhere in the pack, and generation models download into ComfyUI/models/ per the workflow you run.

Gotchas

• Captured output is empty. The captured_image/captured_images/captured_video fields only fill when you actually press Capture/Record in the viewport - there's no auto-capture on Run.
• Viewport aspect doesn't match your video. Set width/height before composing the shot; changing it later re-frames everything.