GNM Head Viewport (Advanced)

The basic viewport gambles; this one sculpts. GNMHeadViewportAdvanced takes the all-in-one head and adds the per-component sliders from GNM's official demo - the same curated subset Google's own notebook exposes - so instead of re-rolling seeds until a face happens to work, you build it.

The component panels cover that demo set: identity gets head ×10 and eyes ×3; expression splits into left eye ×3, right eye ×3, lower face ×7, tongue ×4, and pupil ×1. On top of that it adds the good stuff the basic node lacks - separate neck (±90°) and head (±45°) rotation so you can turn the neck without swinging the head, a full gaze rig (gaze_pitch, gaze_yaw, and gaze_vergence for converging the eyes), and XYZ translate sliders (±0.2m) to frame the head without touching the camera.

Here's the workflow detail that trips people up: the strengths default to 0, not 1. That's deliberate - you start from the template (average) face and sculpt, exactly like the official demo. If you'd rather fine-tune a seeded face, raise identity_strength/expression_strength and the component sliders become offsets layered on top of the sampled vector rather than the whole story.

Under the hood, component values live in a hidden components JSON widget, which is what makes them survive a workflow save/load intact - close ComfyUI, come back tomorrow, your sculpt is still there. Each group also has a reset button for that region only, which saves you from starting over when one area goes wrong. The README even hands you a working example: to open the mouth, set lower-face e200 = -2.6, e201 = 1.0, e202 = -2.1.

Outputs are the standard four - image, normal_map, depth_map, mask - and the live preview behaves exactly like the basic node's: drag to orbit, wheel to zoom, queued render at full width/height. Reach for this one when you want the demo's hands-on sculpting without committing to the full component dump. The Pro node below it is for the truly obsessive.

Install

ComfyUI Manager is the easy path: Custom Nodes → search "ComfyUI-GNM" → Install, then restart. Or clone by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/soylab-edu/ComfyUI-GNM
cd ComfyUI-GNM
pip install -r requirements.txt

Model data ships in vendor/; no separate download, no TensorFlow. Renderer deps (pyrender, pyglet<2, pinned PyOpenGL==3.1.10) and PYOPENGL_PLATFORM=egl on headless Linux, same as the rest of the pack.

Common issues

The one thing that confuses everyone is the strength defaults of 0 - if the face isn't moving when you drag component sliders, check that you actually changed a slider, and that you're not staring at a component whose offset is being dominated by a large identity_strength. And if the component panel never appears, it's the same restart fix as the basic viewport: the /gnm/viewport/* routes only register at startup.