CRM Modeler (Cuda only)

If you've got a CUDA GPU, this is the modeler you actually want. The regular CRM Modeler gets you a vertex-colored .obj - fine for a quick look, embarrassing for anything you plan to put in a scene. The Cuda variant does the full production pass: it cleans up the raw geometry, unwraps UVs, bakes a proper 1024×1024 texture map, and hands you a .glb that game engines and Blender actually like.

What's different under the hood

Same inputs, same output type, same triplane reconstruction: crm_model (a CRM_MODEL from the Model Loader), poses (the six RGB views from the Pose Sampler), and coordinates (the six CCM maps from the CCM Sampler). The difference is in the export path. Where the CPU modeler just exports what the decoder gave it, this node runs the raw mesh through a cleanup pass (via kiui's clean_mesh), then uses nvdiffrast - NVIDIA's differentiable rasterizer - plus xatlas for UV unwrapping to bake the view texture onto the mesh and export it as a .glb.

That's the real reason it's named "Cuda only": nvdiffrast does CUDA-side rasterization. No CUDA, no node. If you try it on a CPU-only box it doesn't fall back gracefully - it raises an exception telling you to use the regular modeler instead.

The one thing to keep in mind: it's slower than the CPU version per frame because of that UV bake, and it needs the CUDA extras installed. But the output is categorically nicer - cleaned geometry, real UVs, a texture that survives exporting instead of a color-per-vertex trick.

Installation - the fiddly part

Pack install is standard (ComfyUI Manager → "ComfyUI-Flowty-CRM", or clone into custom_nodes), but this node is where the Flowty CRM install stops being one-liner territory:

cd ComfyUI/custom_nodes
git clone https://github.com/flowtyone/ComfyUI-Flowty-CRM.git
cd ComfyUI-Flowty-CRM
pip install -r requirements.txt
pip install -r requirements-cuda.txt

requirements-cuda.txt is where nvdiffrast, pygltflib, kiui, xatlas and pymeshlab come from. nvdiffrast is the trap. On Windows it often needs the official build instructions followed by hand, plus a CUDA_HOME environment variable pointing at your CUDA install - otherwise you get CUDA_HOME environment variable is not set and the build dies. If nvdiffrast refuses to cooperate, the README's advice is honest: fall back to the CPU workflow (restart ComfyUI with the CPU-only option) and accept that it's much slower.

Common issues

• CUDA_HOME environment variable is not set - set it to your CUDA toolkit path (see this) and retry the pip install.
• "Cuda is not supported, use the regular CRMModeler node" - you're on CPU; switch nodes or machines.
• Device mismatch errors - same rule as the rest of the pack: keep everything on one device, don't mix a CPU restart with CUDA nodes.
• VRAM - needs the 8GB/16GB floor the README calls out; under 16GB VRAM, run the pipeline as the four split low-vram/ workflows rather than one graph.

You'll need CRM.pth in ComfyUI/models/checkpoints (from huggingface.co/Zhengyi/CRM) - the loader and modeler share it. The payoff is worth the install pain once, and then it just works.