EA Power LoRA

Every serious workflow eventually has four, five, six LoRAs stacked - a style LoRA, a character, a detail enhancer - and the stock way of doing that is a daisy chain of LoraLoader nodes that eats half your canvas. EA Power LoRA collapses the whole stack into a single node with a row-based panel: hit "+ Add LoRA", pick a file, set a weight, repeat. It applies all of them to the model only - no CLIP path - which is exactly what WAN 2.2 UNet-only branches want.

This is the pack's flagship, and the reason it exists: the README's whole framing is "compact LoRA stacking that works cleanly with WAN 2.2 and SD/SDXL/Flux."

How it works

The node takes a model input and a hidden loras_json widget - that JSON is the serialized row list, but you never type it. The web extension gives you the actual UI: checkboxes to enable/disable rows, a dropdown that lists files from ComfyUI/models/loras, a weight field, and remove buttons. Every click re-serializes the rows into loras_json.

At execution it parses the rows, filters out anything not present in your loras folder, and chains ComfyUI's own core LoraLoader per row. Two behaviors worth internalizing:

• Missing LoRAs are silently ignored. Typo a filename and the run just proceeds without it. Good for shared workflows, bad for debugging - check your row names if a style "isn't working."
• Order doesn't change the result. The README is explicit: LoRA application is additive here, so reordering rows is cosmetic. That's a nice property when you're organizing a panel, but don't expect stack-order tricks (like block-merging) from it.

The inputs and outputs that matter

The only required input is model; loras_json is optional and normally hidden behind the UI. The single output is model, ready to wire into your sampler. Per-row, the two fields you'll actually set are the LoRA file and the weight (model strength, two decimals). Weights below ~0 (negative LoRA suppression) work the same as anywhere else.

Because there's no CLIP output, use this node when your pipeline has no text-encoder path - the classic case is a WAN 2.2 graph where conditioning flows through the wrapper rather than a CLIP loader. If you need text-encoder weights too, use EA Power LoRA +CLIP instead; this one is the model-only variant on purpose.

Performance and tuning notes

Each LoRA adds memory and compute. Stacking ten is tempting and cheap to set up, but the README's own advice is to reduce precision or VRAM usage elsewhere if the graph starts chugging. And trigger words: many style LoRAs work fine with CLIP low or off (which is what this node implies), but character/concept LoRAs often want some CLIP weight - and this node gives you none, so for those you want the +CLIP sibling.

Install

Part of comfyui-ea-nodes: ComfyUI Manager → search comfyui-ea-nodes, or git clone https://github.com/ExoticArts/comfyui-ea-nodes into custom_nodes/, restart, hard-refresh. The pack declares no Python dependencies (requirements.txt is empty) and the LoRA list comes straight from ComfyUI/models/loras - no downloads.