LoadStableDelightModel

Most ComfyUI loaders grab their weights for you. This one won't, and that's exactly what trips everyone up. LoadStableDelightModel is the front half of the two-node StableDelight pack (by lldacing, wrapping the Stable-X/StableDelight research release), and it loads the yoso-delight-v0-4-base model that strips specular reflections off textured surfaces. It's a niche task - killing the glare on a glossy product shot, the window reflection in a photo, the hot spot on a car hood - but it's real, and StableDelight is one of the only things that does it. It's your gatekeeper before ApplyStableDelight gets to work.

How it works

The node never calls the Hugging Face API. When the graph loads, it walks every folder inside ComfyUI/models/diffusers looking for a model_index.json, and whatever diffusers-style pipelines it finds become the model dropdown. Pick one and it builds a YosoDelightPipeline from that folder with local_files_only=True - which is Python for "if it's not already on disk, crash instead of downloading."

That local-first behavior is the one thing to internalize about this pack. Most ComfyUI loaders will happily fetch missing weights from the hub; this one assumes you've done the download yourself. You haven't, until you do.

The loader also disables the safety checker, requests the fp16 variant of the model, loads it in half precision on CUDA, and swaps in xformers memory-efficient attention if ComfyUI has it compiled. It's a thoughtful little loader once the model is in place.

The inputs and output that matter

There are only two inputs, and you'll mostly leave them alone:

• model - the dropdown. It's auto-populated from your models/diffusers folder, so if it's empty, your model isn't where the node is looking.
• device - AUTO or CPU. AUTO uses ComfyUI's current torch device. CPU is the escape hatch for machines without a usable GPU, and it's worth knowing it falls back to fp32 there, so expect it to be slow.

The single output, model (DelightMODEL), feeds straight into ApplyStableDelight. There's nothing else to wire.

Installing the model (the part that bites)

Install the pack normally - ComfyUI Manager (search "StableDelight") or:

cd ComfyUI/custom_nodes
git clone https://github.com/lldacing/ComfyUI_StableDelight_ll.git
# restart ComfyUI

Then the model. Note it's a diffusers folder, not a .safetensors checkpoint - it goes in models/diffusers, not models/checkpoints:

cd ComfyUI/models/diffusers
huggingface-cli download Stable-X/yoso-delight-v0-4-base --local-dir Stable-X--yoso-delight-v0-4-base

Or just download the repo contents from Hugging Face into a folder named exactly Stable-X--yoso-delight-v0-4-base. The folder name and the model_index.json inside are what the loader's scan picks up on. The pack itself declares no heavy pip dependencies - pyproject.toml lists only numpy, and everything else (diffusers, transformers, torch) is what ComfyUI already ships.

Troubleshooting

• The model dropdown is empty. You put the files somewhere the scan doesn't check, the folder isn't named right, or the download didn't include model_index.json. Redownload into models/diffusers/Stable-X--yoso-delight-v0-4-base and restart.
• An error about local_files_only or a missing model_index.json on load. Same root cause - the files aren't fully on disk. This is the pack's signature failure, and it's not a bug, it's the design.
• Everything works but it's glacial on CPU. Expected - it's fp32 there. Run it on the GPU.

One honest caveat before you build a whole workflow on it: the model author (both original and this wrapper) note that objects change noticeably, especially semi-transparent ones like lightbulbs and glass. StableDelight is impressive for what it is, but it's a one-shot research model, not a polished tool. Manage expectations, then enjoy the glare-free photos.