Hy3D BPT

Hy3DBPT is the most optional node in this pack, and kijai wants you to know it. It's a wrapper around BPT (github.com/whaohan/bpt), a learning-based remesher that takes your mesh and regenerates its topology with a neural transformer. Where instant-meshes does geometric remeshing, BPT is a generative pass - it resamples your mesh as a point cloud, then has a transformer reconstruct a cleaner, more regular mesh from that cloud.

The readme's framing is worth quoting in full because it's the honest version: "This has some hefty requirements, fully optional, install at your own discretion." We're talking torch_geometric, deepspeed, x-transformers, meshgpt_pytorch, and friends - the kind of dependency stack that can take an evening. This is not a node you install on a whim.

Inputs

• trimesh - the mesh to upgrade.
• enable_bpt (default true) - the kill switch. Set it false and the node passes the mesh through untouched. This is how you keep the node in a shared workflow without dragging everyone through the install.
• seed (default 42) - reproducibility for the generative pass.
• temperature (default 0.5) - sampling temperature for the transformer. Lower = more conservative, closer to the input; higher = more "creative" topology.
• pc_num (default 4096, 1024–8192) - how many points the mesh is resampled into before regeneration. The resolution knob of the whole operation.
• samples (default 100000) - how many candidate samples are drawn from the surface to build that point cloud.

Output: trimesh (TRIMESH).

Should you bother?

Be realistic about what it buys you. BPT's pitch is better topology - a cleaner, more regular structure than raw decode, aimed at people who want to deform or animate the result. That's the same job as Hy3DIMRemesh, done with a much heavier tool. The honest difference is that BPT is a research-grade method (the model name bpt-8-16-500m.pt gives away the ~500M parameter transformer) that can produce noticeably more structured results on organic shapes - at the cost of an install that can genuinely eat an evening and a runtime that's on GPU.

If you just want "clean enough to rig," instant-meshes gets you 80% there for a fraction of the pain. BPT is for the person who specifically wants the neural remeshing and has the patience for the stack. That's a small crowd, and the node is honest about being for them.

Install (only if you mean it)

cd ComfyUI/custom_nodes/ComfyUI-Hunyuan3DWrapper/hy3dgen/shapegen/bpt
pip install -r requirements.txt

Then download bpt-8-16-500m.pt from whaohan/bpt on Hugging Face and drop it into that same bpt folder (that's where the node looks for it). The portable-ComfyUI version of that command is in the pack README. Skip this entirely and the rest of the pack works fine - that's the design.