Hunyuan3D in ComfyUI: distinguish shape generation from texturing
A generated mesh and a textured render are different deliverables. Hunyuan3D-2 separates shape generation from texture synthesis, so a ComfyUI integration must be assessed at both stages rather than described as an ordinary image sampler that happens to save a 3D file.
Start with the actual project
The Tencent Hunyuan3D-2 repository documents a shape component and a texture component. Its examples and model variants are the primary references for the pipeline. A third-party ComfyUI wrapper may expose only some of those capabilities or depend on a particular revision.
Record the wrapper repository, revision, model variant and installation requirements. Do not invent a node identifier from the project name. Use the integration maintainer's complete example and confirm its dependencies before expecting a saved workflow from another machine to load correctly.
Prepare an interpretable input
Use an image you are entitled to process, with a clearly visible object and as little irrelevant background as practical. A single view cannot provide direct observations of every hidden surface. Treat the generated rear and underside as model-generated hypotheses, not recovered measurements.
Preserve the original input and any mask or crop separately. If preprocessing changes the silhouette, record it. Compare multiple attempts with the same prepared input before deciding that a geometry error is caused by the wrapper rather than by ambiguity in the image.
Inspect the mesh before the texture
Export the geometry and open it in the application that will consume it. Check silhouette from several views, disconnected parts, surface defects, orientation and scale conventions. A flattering front-view render can hide defects that matter for animation, printing or a game asset.
Evaluate texture after the geometry passes the intended use case. Inspect seams, projection errors and surfaces not shown in the input. Keep the geometry and texture settings with their respective output files so a failed texture stage does not require repeating every earlier step unnecessarily.
Publish a bounded result
Record the complete environment, model and integration revisions, input file and exported asset. Distinguish an upstream demonstration from a result you reproduced locally. Do not transfer an upstream memory figure to a different variant, wrapper or texture configuration without measurement.
This revision removes the earlier invented KSampler/VAE route and unverified custom-node claims. It provides an asset-review process, not a claim that any photograph becomes production-ready geometry. The acceptance criterion belongs to the downstream task: a visual concept, an editable mesh and a fabrication-ready object require different evidence.