GNM Head Render

If you've ever wanted the exact geometry of a face before the diffusion model invents one, GNMHeadRender is the node you came for. It's the workhorse of ComfyUI-GNM: feed it an identity vector and an expression vector, pose the head any way you like, and out the other side come four things - the color image, a normal map, a depth map, and a mask. The normal and depth maps are the point, because they're exactly what ControlNet-style conditioning eats for breakfast.

The context matters here. ControlNet conditions generation on spatial structure - edges, depth, pose - so the prompt decides what appears and the condition decides where it goes. Most people extract that structure from a photograph with a preprocessor. GNM is the weirder, more powerful path: generate the structure yourself, from a head you can pose and sculpt, so the condition is exactly the face you designed instead of the face a preprocessor happened to find. It's the same impulse as the "consistent characters via geometry" crowd - you're not hoping the prompt holds, you're handing the sampler a map.

How it works, mechanically: the model runs a forward pass - gnm(identity, expression, rotations) - to produce the mesh vertices, then the pack's own renderer (it drives pyrender directly, deliberately bypassing GNM's osmesa-forcing wrapper that breaks on Windows) draws the color, normal, and depth passes plus a mask. The heavy hitters among the inputs:

• gnm_model - from the loader. Required.
• identity and expression - optional. Leave them unwired and you get the average/template face, which is a feature, not a bug.
• head_yaw/pitch/roll - neck-joint rotation in degrees; eyes_yaw/pitch - eyeball rotation.
• camera_azimuth/elevation - orbit the camera (0/0 is dead-on front), with zoom and fov for framing.
• surface - plain (flat material, cleanest normal/depth) or edgeflow_texture (GNM's default texture for the color pass).
• mesh_parts - face_default (eyelid skin excluded, the GNM default) or all.
• width/height - up to 4096; antialias 1–4; background gray level 0–1.

Outputs: image, normal_map, depth_map, mask. Wire the maps into a depth or normal ControlNet Apply and render away. One caveat the ecosystem keeps rediscovering: the post-Flux union ControlNets trimmed their condition menu and normal maps were among the casualties - so this shines brightest on SD/SDXL, or wherever a normal ControlNet still exists.

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

The model data ships in vendor/ (a 51MB npz) - no separate download. The renderer needs pyrender, pyglet<2, and the pack's pinned PyOpenGL==3.1.10 (pyrender's default 3.1.0 breaks on Python 3.13 / numpy 2).

Common issues

The render node is where this pack's environment quirks live. Headless Linux servers need PYOPENGL_PLATFORM=egl set before launch - the README calls it out explicitly; Windows should just work via WGL. If you hit a pyrender import error, the requirements didn't land, and mind the pyglet<2 pin because newer pyglet breaks pyrender's GL context. Expect the first render to be slow while the model loads and caches; everything after that is fast.