Hy3DDelightImage

Photos carry light the way bags carry germs. A highlight on the nose, a hard shadow under the chin - pass that straight into Hunyuan3D's shape model and the geometry comes back with those baked in, giving you a 3D model with lighting sculpted into the surface. Hy3DDelightImage is the wash-your-hands step: it runs your input image through the delight model (DownloadAndLoadHy3DDelightModel) and returns a de-lit version, ready for the shape generator.

Mechanically it's an InstructPix2Pix-style edit with an empty prompt - the model is trained to remove illumination, so it doesn't need to be told what to do. The node upscales your image to the target width/height with Lanczos, runs the pipeline, and hands back a de-lit IMAGE.

Inputs that matter

• delight_pipe - from the delight model loader. Required, obviously.
• image - your input photo or render.
• steps (default 50) - diffusion steps. 50 is the standard here; you can drop it for speed when iterating.
• width / height (default 512×512) - the working resolution. The node upscales to this before running, so match it to your source roughly; there's no benefit to a huge canvas here since the model is 512-class.
• cfg_image (default 1.0) - image guidance strength. This is the one dial worth touching: it controls how hard the model holds onto the original image while removing light. Higher sticks closer to the source (and leaves more lighting); lower strips more but can wash out fine detail. 1.0 is the sensible starting point.
• seed (default 42) - for reproducibility.
• scheduler (optional) - a NOISESCHEDULER from Hy3DDiffusersSchedulerConfig, if you want to swap the default Euler A.

Output is a single de-lit image, which feeds Hy3DGenerateMesh.

When to skip it

If your input is already a clean, flatly-lit render from a 2D model, the delight stage is optional - the shape model handles it fine and you save a full diffusion pass. The delight branch earns its keep with real photographs: phone shots, product photos, anything with visible lighting. A common middle path is to run it once, keep the seed, and use the result as your reference image for the whole session.

Notes

It's pure diffusers under the hood, so it's one of the friendlier nodes to run - no compiled rasterizer involved. Install the pack (ComfyUI Manager, search "Hunyuan3DWrapper", or clone into custom_nodes and pip install -r requirements.txt), let the delight model download once into models/diffusers/, and you're set. If the first run errors on a partial download, delete the half-downloaded folder and re-run.