ModelScope-LoRA 多LoRA加载

The ModelScope image and edit nodes each have three LoRA slots, and filling all three with raw model IDs is tedious. This node is the multi-lane version of the pack's single LoRA loader: three preset dropdowns, three weights, and six outputs that plug straight into those slots. Nothing fancy, nothing hidden - if you're stacking a style LoRA with a character one and a detail fix, this is the wiring hub.

How it works

Like its single-loader sibling, it makes no API call. It reads the lora_presets list from modelscope_config.json and resolves three selections into IDs and weights.

The inputs

The inputs are three lora_preset dropdowns (lora1_preset, lora2_preset, lora3_preset), each with the same options - the five bundled Chinese-style presets plus 无LoRA (no LoRA). Below those sit per-lane weights and toggles:

• lora1_weight / lora2_weight / lora3_weight (0–2, default 0.8) - the weights you want, when you want custom ones.
• lora1_use_custom / lora2_use_custom / lora3_use_custom (booleans, default off) - per-lane switch between "use the preset's stored weight" (off) and "use the weight I typed" (on). You can mix: preset weight on lane one, custom weight on lanes two and three.

Wiring it up

Outputs are six values that match the image node's slot names exactly: lora1_id, lora2_id, lora3_id (STRINGs) and lora1_w, lora2_w, lora3_w (FLOATs). The naming lines up one-to-one with the lora1_id/lora1_w … lora3_id/lora3_w inputs on ModelScopeImageNode and ModelScopeImageEditNode, so wiring is literally dragging each output onto the matching input. All six get packed into the API request's loras dict when the image node runs; the first LoRA also gets promoted to the request's primary lora/lora_weight fields.

Two things worth knowing

Two things worth knowing. First, a preset set to 无LoRA resolves to an empty ID, so it's fine to leave lanes unused - the image node skips empty ones. Second, the dropdown options are whatever's in modelscope_config.json at node-creation time; add presets via the ModelScopeLoraPresetNode and you'll see them on fresh nodes (or after reloading the workflow), not on already-placed ones.

Installing it

Install is the standard pack routine - ComfyUI Manager, searching "ComfyUI ModelScope API Node", or:

cd ComfyUI/custom_nodes
git clone https://github.com/hujuying/ComfyUI-ModelScope-API

then restart ComfyUI. No extra dependencies; this is pure data plumbing, same as the single loader. Use this node when you stack two or three LoRAs, and the single-loader when you only ever use one - the outputs are identical, so swapping between them is a re-wire, not a re-think.