☁️BizyAir Hy3D_2_1SimpleMeshGen

Hunyuan3D is Tencent's line of open 3D-asset generation models, and it occupies genuinely uncontested ground - the KB's read is that "Hunyuan3D and HunyuanWorld have no Apache 2.0 equivalent from Alibaba," which is unusual in an ecosystem where almost every capability gets a rival within weeks of release. This node runs the 2.1 shape-generation model specifically - the model dropdown names it directly - turning a single image into a raw 3D mesh.

How it works

Feed it one image and it infers 3D geometry from that single 2D view - no other angles, no reference turnaround. This is the "simple" variant: shape only, no texture baked in. If you need a textured result, you're looking for a fuller Hunyuan3D pipeline elsewhere in the pack, not this node.

The inputs and outputs that matter

• model - an enum with one current choice, hunyuan3d-dit-v2-1/model.fp16.ckpt, naming the 2.1 shape DiT directly.
• image - required; the node's own tooltip is straightforward: "Image to generate mesh from." Works best on a clean single-subject shot with a clear silhouette - a product photo or character render rather than a busy scene - since the model has only one view to infer geometry from.
• steps (default 50, 1-100) - "Number of diffusion steps," per the tooltip. More steps costs generation time for (presumably) cleaner geometry, the same tradeoff as image diffusion step counts.
• guidance_scale (default 5, 1-30, step 0.1) - "Guidance scale," per the tooltip. Same conceptual role as guidance/CFG elsewhere: how strictly the generated shape sticks to what the image implies versus the model's own priors.
• octree_resolution (default 384, 32-1024, step 32) - "Octree resolution," per the tooltip. Controls the resolution of the octree structure the mesh gets extracted from. Higher means finer geometric detail, but a heavier mesh and slower generation - 384 is a reasonable middle ground; push toward 512-1024 only if you actually need the extra fidelity, since both mesh size and generation time climb with it.

Output: trimesh (TRIMESH) - a raw mesh object, ready for export or a downstream texturing/rendering node.

How to install it

ComfyUI Manager: search BizyAir, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir

then restart. BizyAir wants an account and API key set up on first use per the README - this is Tencent's model wrapped for BizyAir's execution graph, not a local model file you manage yourself.

Common issues & troubleshooting

Geometry looks wrong or incomplete on the far side of the object. Expected limitation of single-image 3D generation in general - the model has no information about what it can't see, so occluded surfaces get inferred rather than reconstructed. Cleaner, more front-facing source images with an unambiguous silhouette give the most reliable results.

Mesh is too heavy or generation is slow. Lower octree_resolution before touching anything else - it's the parameter most directly tied to both output size and render time.

Need a textured result, not bare geometry. This node is explicitly the "simple" shape-only variant - look for a fuller Hunyuan3D texturing node elsewhere in the pack rather than expecting this one to produce it.