(Down)Load Hy3D DelightModel

If you feed Hunyuan3D a photo with studio lighting or a harsh sun, the 3D model comes back with baked-in highlights and shadows painted onto the mesh. That's what the delight stage exists to fix: DownloadAndLoadHy3DDelightModel loads the model that strips lighting out of your input image before the shape model ever sees it. It's the "delight" in "de-light" - removing the light, not making you happier (though it does do that too).

The name is doing honest work: this node downloads and loads the hunyuan3d-delight-v2-0 model from Tencent's Hugging Face repo, then exposes it as a delight_pipe you feed into Hy3DDelightImage. It's a full diffusers pipeline under the hood - specifically a StableDiffusionInstructPix2PixPipeline - which is a cute detail: Tencent repurposed the "edit this image according to instructions" architecture with an empty prompt, so the model just removes lighting without being told to.

What it does, step by step

• Downloads hunyuan3d-delight-v2-0 from tencent/Hunyuan3D-2 into ComfyUI/models/diffusers/ on first run. Yes, it's a multi-gigabyte download the first time. That's normal.
• Loads it as fp16 with the Euler Ancestral scheduler and enables model CPU offload (unless you connect compile_args).

The only required input is the model dropdown, which currently has exactly one choice (hunyuan3d-delight-v2-0), so don't overthink it. The optional compile_args socket is a HY3DCOMPILEARGS from the pack's Hy3DTorchCompileSettings node - torch.compile of the UNet/VAE for speed. The tooltip is worth quoting: it requires Triton, and torch 2.5.0 is recommended. If you don't have Triton set up, leave it unconnected and let CPU offload do its thing.

Output is a single delight_pipe (HY3DDIFFUSERSPIPE) → Hy3DDelightImage.

Where it sits in the workflow

LoadImage → ImageRemoveBackground → Hy3DDelightImage → Hy3DGenerateMesh. The delight stage is what lets you use a real photo as your input instead of a flat AI-rendered image. If your input is already a clean render with no lighting baked in, you can honestly skip the whole delight branch - many workflows do and the shape model is fine with it.

Gotchas

One real-world complaint that keeps showing up: the download. The node pulls from Hugging Face, and if the snapshot download hiccups (network, disk space, HF rate limits) the node errors out on a partially-populated folder. Re-running the queue triggers it to re-attempt. And the usual pack-wide warning applies - this pipeline loads fine on its own, but it's the paint stage later that needs the compiled custom_rasterizer wheel on Windows (Python 3.12 / CUDA 12.6). The delight model itself is plain diffusers, so it's the friendliest part of the install.

Install: ComfyUI Manager, search "Hunyuan3DWrapper", or clone into custom_nodes and pip install -r requirements.txt. Then just run the node and let it download.