LoRA Bridge Loader

LoRA Bridge Loader is the front door to PG Nodes' whole LoRA toolkit. Instead of four separate Load LoRA nodes sitting in your graph, it loads up to four LoRA files once and packs them into a single LORA_BRIDGE bundle. That bundle then feeds either the hook-based (non-destructive) path or the merge/bake path - the rest of the PG LoRA family decides what happens to it.

If you've fought with a ComfyUI graph that has six LoRA loaders cross-wired into model and clip, you get the appeal instantly: one node, four dropdowns, one wire out.

How it works

Four inputs - lora1 through lora4 - each a dropdown over your loras folder, defaulting to none. It loads whichever you pick via ComfyUI's own load_lora, then hands you a single LORA_BRIDGE output containing the loaded files plus the device they landed on. Because LoRA weights patch both the UNet and CLIP paths, the bridge holds the same object for both - you don't load anything twice.

A device_cpu boolean toggle moves the loaded weights to CPU. It's "best-effort," like the pack's other CPU switches: useful on low-VRAM rigs where you'd rather eat a slower load than blow the budget, but not something to treat as a guarantee.

From the bridge you have two wiring styles:

• Hooks (non-destructive): bridge → Create LoRA Clip Hook / Model Hook → (optionally Set Hook Keyframes) → Set Clip/Model Hooks. The LoRA patches the pipeline live at inference; you can keyframe its strength over the denoising run and toggle it without reloading.
• Merge (bake-in): bridge → Merge LoRA Model / Clip / Both. Weights get baked into the model or clip permanently for that run - faster at runtime, but fixed strength and no switching mid-run.

The inputs that matter

• lora1–lora4 - pick up to four files. Leave extras on none; they're skipped cleanly.
• device_cpu - offload loaded weights to CPU. Off unless you're scraping VRAM.

Output is the single LORA_BRIDGE bundle - there's nothing else to wire, which is the whole design.

Install

Comes with the PG Nodes pack:

ComfyUI Manager → search "PG Nodes" → Install → restart

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/GizmoR13/PG-Nodes   # rename folder to PG-nodes

Restart ComfyUI and it appears under PG/Lora. No extra dependencies; it leans on comfy.hooks, which is part of modern ComfyUI. If you're on a very old ComfyUI build, the hook nodes error out telling you the hooks API is missing - update ComfyUI first.

Common issues

• Hook does nothing. If a bridge slot is empty or the file failed to load, the hook nodes pass through silently rather than erroring. The console shows [LoRA Bridge] ERROR: LoRA file invalid or unsupported format. - check the terminal output, not just the graph.
• Strength "does nothing" on one path. Some LoRAs only touch cross-attention, so clip-only vs model-only hooks genuinely differ. Try the other path before assuming a bug.
• New LoRA not in the dropdown. File lists are built at import; restart ComfyUI after dropping a new .safetensors into models/loras.

Worth knowing: this is a newer, self-taught-author pack, not a battle-tested giant like the Impact Pack - but the LoRA bridge is one of the more thoughtfully designed pieces of it, and it shows.