Nodes/ComfyUI_SGLDiffusion_Fix/SGLDiffusion Server Unset LoRA
ComfyUI Node

SGLDiffusion Server Unset LoRA

The off switch for LoRAs you can't see

By endman100·Created 5 months ago·Updated 5 months ago· 0
SGLDiffusion Server Unset LoRA
  • sgld_client
  • sgld_client
targetall

The smallest node in the pack, and the one you'll forget until it bites

SGLDiffusion Server Unset LoRA does exactly one thing: tells your SGLang Diffusion server to drop a loaded LoRA. It's the mirror image of Server Set LoRA, and in the pack's Server Mode it's the difference between clean generations and generations that quietly carry yesterday's style.

Why it exists at all comes down to how Server Mode works. The model and its adapters live on a remote SGLang Diffusion process, and adapters you set stay loaded between requests - there's no "workflow reset" the way there is in a local ComfyUI graph. Set a LoRA for one image and the server keeps applying it to everything after, until something explicitly unloads it. This node is that something.

Inputs and behavior

  • sgld_client - required, from the SGLDiffusion Server Model node. Same wire that feeds Set LoRA and the generate nodes.
  • target - all by default, or transformer, transformer_2, critic. On multi-tower video models this is how you strip a LoRA off just one component instead of everything.

The output is the sgld_client again, so you can chain it: Set LoRA → Generate → Unset LoRA → next thing, all on one wire. That chain is actually the idiomatic way to use these in a graph - set the adapter, generate, then clear it so the next run starts from clean weights.

The honest gotcha

Because it's stateful and the state lives on the server, the failure mode is silent. There's no visible cue in ComfyUI when a LoRA is still loaded; you just notice outputs drifting. The cheap habit is to treat Set/Unset as a matched pair and always clear after you're done. If a run comes out inexplicably off-model, add an Unset node with target all at the top of the branch - it's cheap insurance.

Install

Nothing special here - it's part of the shared pack, and the pack still ships no requirements.txt, so SGLang itself is the manual step:

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

Restart ComfyUI after the clone, and you're set. This node is so thin that there's not much else to say about it - which is exactly how a good off switch should be.

CategorySGLDiffusion

Inputs (2)

NameTypeDefaultDescription
sgld_clientSGLD_CLIENT
targetoptCOMBOallWhich transformer(s) to unset the LoRA from

Outputs (1)

NameTypeDescription
sgld_clientSGLD_CLIENT