LoRA Loader (CLIP)

A plain LoRA Loader patches the diffusion model and ignores the text encoder. Most LoRAs are trained that way and that's fine - a model-only patch covers the vast majority of style and character work. But plenty of LoRAs - character LoRAs trained with the text encoder on, some style LoRAs that lean on trigger phrasing - carry a second set of weights that nudge how your prompt gets embedded, and if you skip those you're running the LoRA at maybe half its intended strength. LoRA Loader (CLIP) from ComfyUI-NynxzNodes is the version that patches both: same one-node stack, CLIP in and out alongside MODEL.

Think of the two loaders as siblings: identical row-based stack UI, identical searchable picker with bookmarks and previews, identical on/off dots per row. The only difference is that this one takes a clip input and returns a clip output in addition to model. If you don't need CLIP patched, the plain loader is the one to use - it's one less socket to wire and one less thing ComfyUI has to hold in memory.

How it works

Same additive patch-application as any LoraLoader chain, just consolidated: each row's LoRA is applied to both the model and the CLIP object in stack order. The per-row strength applies to both patches at once - there's no separate "model strength vs clip strength" split per row here, which is simpler than the A1111-style separate sliders and occasionally the reason a LoRA feels weaker or stronger than you remember from another tool.

The stack itself is a special widget, so rows live on the node: on/off dot, searchable picker over models/loras, preview thumbnails, bookmarks, strength slider with negatives. Bookmarked favorites persist across sessions, which is the feature you won't appreciate until you've got forty LoRAs and three you actually use.

Inputs: model, clip, and the stack. Outputs: model and clip, both patched, wired on to your sampler and your prompt encode respectively. Remember the same additive-collision caveat as everywhere: stack two character LoRAs and both deltas fire on every token. If you need separate characters in one image, this isn't the tool - that's what the pack's Regions group is for.

Install

Part of Nynxz's Custom Nodes: ComfyUI Manager → search "NynxzNodes" and install from the registry, or

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-NynxzNodes

Zero pip dependencies. Clone-by-hand users: the pack's Vue UI must be built before the pickers render (pnpm install && pnpm build; web/ is gitignored), so the registry/Manager route is the easy one. Restart ComfyUI, and update ComfyUI if the node list looks wrong - the pack uses the newer extension API.

Gotchas

If your favorite LoRA behaves identically whether CLIP is patched or not, that's normal - it may simply carry no text-encoder weights, and the CLIP loader costs you nothing extra beyond a wire. The real trap is forgetting the separate-strengths thing: a character LoRA that wanted its text side at 0.8 and its model side at 1.0 is now forced to one number, so if a familiar LoRA feels off, suspect that coupling before you suspect the install.