Power Lora Loader (Eden) 🎲

Stack four LoRAs to get one look and you've got four loader nodes chained in a line, each with two strength fields, and a graph that looks like a fuse box. rgthree's Power Lora Loader fixed that years ago and became one of those nodes people install before anything else. Eden's version is a reimplementation of the same idea with one extra trick: any LoRA's strength can be drawn from a range by a seed, so a workflow explores a strength space instead of one fixed point in it.

How it works

The node draws its own rows in the graph: hit ➕ Add Lora, pick from the list ComfyUI serves at /models/loras, and you get a row with a toggle, a name and a strength. Click the 🎲 on a row and a second line opens where you set min and max. There's a Toggle All switch at the top, and right-clicking a row gives you Move Up and Remove.

Each row serializes into the prompt as a lora_N entry; the Python side reads them in order and calls ComfyUI's own LoraLoader per row, passing the same strength to model and clip. When 🎲 is on, the value is sampled uniformly between min and max - and the random stream isn't shared, it's derived per row from the seed plus the LoRA's file name. That's the detail that makes this usable: add a sixth LoRA, delete the second, reorder the rest, and every other row keeps the strength it had.

The inputs that matter

• seed - seeds the strength sampling and nothing else. It has the usual control_after_generate dropdown, so leaving it on randomize gives every image in a queued batch fresh strengths. Queue twenty, keep the ten that landed.
• model - the diffusion model to apply LoRAs to. Optional, and usefully so: with nothing connected the node just samples strengths and returns them, a lottery you can feed elsewhere.
• clip - gets the same LoRA at the same strength as the model. Note the limit in that sentence: no separate clip strength. If a LoRA wants 0.6 on the clip and 1.0 on the UNet, this node can't express it and core LoraLoader still can.

The outputs

MODEL and CLIP go where a normal loader's outputs go. The interesting ones are settings and strength_1 … strength_N.

settings is a JSON blob - the seed plus each LoRA's applied strength keyed by file basename, {"seed": 3, "loras": {"style_a": 0.72}}, disabled rows at 0.0. Run it into Save Param Dict (same pack) and every render logs what the dice rolled, which is the difference between a lucky folder of images and a dataset you can reason about later.

The strength_N outputs are one FLOAT per row, named after the LoRA in the UI - wire a sampled strength into a Math node or another loader if you want to use it elsewhere. Remove a row and the outputs compact up, so double-check links after you prune.

Installing it

You install the pack, not the node. In ComfyUI Manager search Eden Comfy Pack, or:

comfy node install eden_comfy_pipelines

Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/edenartlab/eden_comfy_pipelines.git
pip install -r eden_comfy_pipelines/requirements.txt

Restart, then look under Eden 🌱 → Loaders. The requirements are light - scikit-learn, scipy, opencv-python, py7zr, no torch of its own - and you need ComfyUI 0.3.48+. This node wants no API key and no network.

Where people get burned

• A new row defaults to 🎲 on, range 0.5–1. Add a LoRA, don't look, and your intended 1.0 is actually 0.73. Turn the dice off for anything you want fixed.
• "Off" is not "bypassed." The toggle zeroes the strength, but with a model connected the loader still runs LoraLoader on that file at 0.0. No visual effect, correct - but you don't get the load time or patch memory back either.
• Sixteen rows is a hard cap. Go over and the node raises rather than quietly clipping.
• Order matters. Put the big structural LoRAs earlier and the gentle ones later; when something looks washed out, test the order before blaming the strengths.
• It's a canvas-drawn custom widget. Eden's frontend paints the rows itself, the same family as rgthree's loader - and that one took visible damage from ComfyUI's Nodes 2.0 frontend. If the rows render as a smear after an update, try the legacy canvas before rebuilding the workflow.

And the one that isn't this node's fault: if a LoRA does nothing, check it matches the base architecture, that its trigger word is in the prompt, and that Pony/Illustrious anime LoRAs often want CLIP skip 2.