SGLDiffusion Server Set LoRA

The LoRA switch that lives on the other machine

SGLDiffusion Server Set LoRA is how you bolt an adapter onto a model that's being served by SGLang Diffusion - which is to say, a model that lives somewhere your normal ComfyUI LoRA workflow can't reach. In Server Mode the sampling happens on a separate box, so the usual "pick a file from ComfyUI/models/loras" trick doesn't apply. This node tells the server to load a LoRA it can see, and it's one half of a pair: Server Unset LoRA is the off switch.

It's a pass-through node in the mechanical sense - the only output is the same sgld_client you fed in - but don't underestimate it. Server-side LoRAs are stateful: once you set one, it stays loaded and keeps affecting every subsequent generation until you explicitly unset it. If your second image mysteriously carries style from your first, you forgot this.

The inputs that matter

• sgld_client - from the SGLDiffusion Server Model node. The only required wire.
• lora_name - and here's the gotcha. Despite the name, this is not a file in your loras folder. It's a path or name the server can resolve - the adapter file has to exist on the machine running SGLang. If you've been uploading LoRAs to your local ComfyUI, that's not the same place.
• lora_nickname - optional. Leave it blank and the node uses the LoRA's filename (without extension) as its nickname. Set it if you want a stable handle, especially since the Set/Unset pair match on target rather than nickname.
• target - all by default, or transformer, transformer_2, critic. On single-tower models (FLUX, Z-Image, Qwen-Image) you can leave it alone. On multi-tower video architectures that separate the denoiser from the text critic, this is how you apply a LoRA to just one of them. If you don't know what those towers are, all is the right answer.

How it works

Under the hood the node is a one-line call to the server's LoRA endpoint: it sends nickname, path, and target, and the server hot-loads the adapter without restarting. That hot-swap is the whole appeal - in Server Mode you're tuning a remote inference service, and this is the knob.

Install and caveats

Shared pack install - clone into custom_nodes (ComfyUI Manager, search "SGLDiffusion"), restart, and make sure SGLang itself is installed since the pack doesn't ship it:

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

Realistic troubleshooting is mostly about paths. Failed to set LoRA adapter on a server that's otherwise responding almost always means the server can't find the file at the lora_name you gave it - so check the server's filesystem, not yours. And remember the stateful behavior: after a failed or unwanted load, hit the Unset node with target all to get back to clean weights rather than stacking adapters you can't see.