SGLDiffusion LoRA Loader

Integrated Mode's familiar-looking LoRA node

If you've used ComfyUI for more than a week, SGLDiffusion LoRA Loader looks like a standard LoRA loader: a MODEL in, a MODEL out, a dropdown for the LoRA. That familiarity is the point - it's the LoRA companion to the pack's Integrated Mode, where the SGLDUNETLoader hands sampling over to SGLang but everything else (CLIP, VAE, conditioning, the rest of your graph) stays ordinary ComfyUI. This node slots into that world like any loader you already know.

The big difference from the server-mode Set LoRA node: here the dropdown actually lists your local loras/ folder. The README's suggested upgrade path is literally "replace the official LoRA loader with this one" in workflows that already use the SGLDUNETLoader, and that swap is painless precisely because the surface looks identical.

How it works - the part that isn't familiar

The mechanism is where it departs from a plain loader. Instead of merging LoRA weights into the model at load time the normal way, this node registers the adapter into the model patcher's patches dict and then calls SGLang's own set_lora - the runtime adapter API, not a weight merge. The same SGLang machinery that hot-loads LoRAs on the server side is doing it in-process here. That's why it can stack multiple adapters and why the nickname field exists at all: nicknames are the keys in that patches dict, and SGLang addresses adapters by them.

The inputs that matter

• model - the MODEL out of SGLDUNETLoader. Required.
• lora_name - a real dropdown of your local loras/ files. The nice, boring part.
• strength_model - 0 to 10, default 1.0. Same semantics as every other LoRA strength slider; on distilled models you'll often want it lower.
• nickname - optional string. Leave blank and the node auto-generates one. Set it only if you have a reason to reference the adapter by name later; most people never touch it.
• target - all, transformer, transformer_2, or critic. Same tower-selection idea as the server node: leave it on all unless you're working a multi-tower video model and know what you're doing.

Gotchas

Because adapters accumulate in the model patcher, swapping LoRAs mid-graph means being deliberate about which loaded model you're feeding the sampler - a stale patcher keeps its list. And remember this node only exists to serve Integrated Mode: it does nothing useful without a model from SGLDUNETLoader, and unlike the server node it can't reach adapters living on a remote machine.

Install

Shared pack install, with SGLang as the real dependency:

cd ComfyUI/custom_nodes
git clone https://github.com/endman100/ComfyUI_SGLDiffusion_Fix
pip install sglang[diffusion]

Restart ComfyUI after the clone. If the dropdown comes up empty, it's the same fix as anywhere else in ComfyUI - the LoRA needs to be in ComfyUI/models/loras/, not wherever it ended up on download.