Multi Lora Loader Advanced

Two LoRAs in one workflow is where the trouble starts. Character plus style, motion plus aesthetic - stacked adapters don't stack cleanly; they fight over the same denoising steps, and your only knob on a stock Load Lora node is one static strength applied to the whole generation. Multi Lora Loader Advanced (from GiusTex's ComfyUI-MoreEfficientSamplers pack, a rename/spinoff of his GTNodes bundle) is the "let me be more specific" answer: it loads several LoRAs at once, and optionally lets you say exactly when each one gets to act - start at step 12, out by step 24, that kind of thing. The name is a lie in a good way: it doesn't train anything, and it doesn't replace your sampler. It just decides when a LoRA's strength is non-zero.

What it actually is

It's a multi-LoRA loader with a scheduling option bolted on. The core mechanism is dead simple: the node's num_loras widget is a dynamic combo (1 to 20 slots) that grows a row of inputs for each LoRA - lora_name, strength_model, start_lora_at_step, and end_lora_at_step. Every slot left on None is skipped. Changing num_loras wipes all the values you set, which is the README's own loud warning: start at 3–4 slots and leave the extras on None.

Where it gets interesting is the optional sigmas input. Connect the same sigmas tensor you're feeding your sampler, and the node hands each LoRA to ComfyUI-SigmaSync-LoRA (capitan01R's node, not this pack). It builds a per-step strength list full of zeros, fills in your LoRA's strength only between its start and end step, and forces SigmaSync's "explicit" curve so nothing clever happens between the anchors. end_at_step of -1 means "keep it until the last step" - the default, and the sane one. Leave sigmas unconnected and the whole scheduling layer quietly vanishes: the node falls back to the plain native LoRA loader and applies each LoRA's full strength from step 0, start/end steps ignored.

What you set

The inputs that matter, roughly in order:

• model - the model you're adding LoRAs to. Required.
• num_loras - the dynamic combo that decides how many slots exist. See the reset warning above.
• sigmas - optional, and the whole point of this node. Wire in the sigmas from your sampler or scheduler (the same ones you pass to SamplerCustomAdvancedEfficient or a Sampler Custom Advanced). Without it, you just have a plain multi-LoRA loader.
• Per-slot strength_model (0–10, default 1.0) and the two step widgets, which only do anything when sigmas is connected.

Output is a single model ("Model with added loras") - wire it into your sampler exactly where the checkpoint's model output would go.

Why you'd bother

Scheduling a LoRA in time is the trick people reach for when two adapters collide: run the character LoRA early to lock composition, then fade the style LoRA in late so it doesn't fight it - the standard trick for combining a style with a character without block editing. It's also genuinely handy on video models where you want a "lightx" or motion LoRA active for only part of the denoise. If your workflow has one LoRA and one strength, this node is overkill; if you're stacking three and cursing, it's the specific tool.

Installing it

ComfyUI Manager is the easy route - search "MoreEfficientSamplers". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/GiusTex/ComfyUI-MoreEfficientSamplers

Then restart ComfyUI. No model files, no requirements.txt, no Python deps to fight. Two real gotchas: the dynamic combo needs ComfyUI 0.8.1+ with frontend 1.33.4+, so update if the widgets won't render; and if you want the scheduling path at all, you must also install ComfyUI-SigmaSync-LoRA or the node throws an import error the moment you connect sigmas.

Where people get burned

• Connect sigmas, forget the dependency - instant import error. SigmaSync is the one thing this node can't run without.
• Set a start/end range that's outside the step count - every entry in the strength list stays zero, and the node silently returns a clone with no LoRA applied. If a LoRA "isn't loading," check your step range before you suspect the file.
• Change num_loras mid-tune and lose an afternoon of strength values. Pick your slot count first, then tune.

One honest caveat: the pack's README still reads like GTNodes' (it tells you to clone ComfyUI-GTNodes). The actual repo is ComfyUI-MoreEfficientSamplers - and it also ships the pack's main attraction, SamplerCustomAdvancedEfficient, whose sigmas output pairs naturally with this node.