Hunyuan3D: Image(s) to Model

Hunyuan is Tencent's in-house model program, and it owns 3D more aggressively than it does video these days - which is why a Tencent image-to-3D node ships in ComfyUI core at all. This is its image front door: one image (or up to four views) in, a model out - and it's the node in this family with the most generous output set: GLB, OBJ, the texture image, and the metallic/normal/roughness maps as separate outputs. If your workflow lives on extracting PBR maps for a game engine, that's a genuine differentiator.

Standard family fine print: paid cloud API via Comfy Org proxy (account + credits, roughly $0.30–$0.70 depending on options), no local compute. Tencent's pricing model also layers on per-feature credits - multiview images, PBR, and a non-default face count each add a slice.

How it works

Upload your image (plus optional image_left, image_right, image_back for multiview consistency) and pick a model. The node sends the job to Tencent's 3d-pro API, polls until it lands, and downloads the results. The OBJ result arrives as a ZIP that the node unzips and splits into the mesh plus texture maps, all surfaced as separate outputs. The images are downscaled server-side before upload (max side ~4,900px), so you don't need to pre-optimize.

The inputs that matter

• model - 3.0 or 3.1. The tooltip carries a hard rule: LowPoly is unavailable for the 3.1 model, so pick 3.1 only for Normal/Geometry generation.
• image - your main reference, at least 128×128.
• face_count - 3,000 to 1,500,000 (default 500,000). Your triangle budget; the price badge bumps if you move off the default.
• generate_type - a dynamic combo with three branches: Normal (adds a pbr toggle), LowPoly (adds polygon_type - triangle/quadrilateral - plus pbr), and Geometry (bare mesh, cheapest).
• seed - the usual non-deterministic re-run control.

Outputs: model_file (STRING), GLB, OBJ, texture_image, and the three optional PBR maps (optional_metallic, optional_normal, optional_roughness) - each as an IMAGE you can route into other nodes.

Common issues

The LowPoly-on-3.1 error is the most common trip-up - it's a real server-side restriction, not a bug. Second, the face-count dial is where credit costs creep up; 500K is the default for a reason. And don't expect the PBR maps to be perfect out of the gate - they're generated, so checking them in a viewer before committing to a pipeline is worth the minute. The multiview inputs help a lot with back-of-model consistency if the front-only result comes back mushy; they cost extra, but they're cheaper than a second full run.