Lora Loader

At its core this is the standard ComfyUI LoRA loader - it patches a LoRA onto both your model and your CLIP so the LoRA's style, character, or concept actually shows up in the output. What art-venture adds are two small conveniences that matter more than they sound: an enabled toggle so you can mute a LoRA without ripping it out of the graph, and a lora_override string so you can swap which LoRA loads without touching the dropdown. If you build workflows you run repeatedly, or drive them by API, those two make life easier.

A quick reminder of what a LoRA even is: a small low-rank adapter, usually 10-200 MB, that patches a checkpoint's attention weights to add something the base model didn't know - a character, an art style, a concept. It's a patch, not a model, so it generates nothing on its own; it rides on top of a compatible checkpoint.

How it works

The node takes your model and clip, applies the named LoRA to both at the strengths you set, and passes the modified model and clip back out. Because it outputs the same two types it takes in, you chain them - one AV_LoraLoader after another to stack several LoRAs. The pack's own description says as much: multiple LoRA nodes can be linked together.

The inputs that matter

• lora_name - the dropdown reads from your models/loras folder. Empty list means you have no LoRAs installed there.
• strength_model - how hard the LoRA hits the diffusion model (default 1). The technical range is a silly -100 to 100; ignore that. Real-world, the old SDXL-era rule was 0.5-0.8 because 1.0 overpowers the prompt - but that advice does not generalize. On Z-Image people run 1.0 and above because dropping it costs likeness, and some LoRAs (PerfectHands) want ~1.6. Check the LoRA's model page instead of reflexively reaching for 0.7.
• strength_clip - how hard it patches the text encoder (default 1). Usually you keep it at or near strength_model; lowering it softens how strongly the LoRA's trigger words pull.
• enabled (optional) - flip it off to bypass the LoRA in place. Great for A/B comparisons.
• lora_override (optional) - pass a LoRA filename by string, handy for automation.

Installing it

ComfyUI Manager → search comfyui-art-venture → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/sipherxyz/comfyui-art-venture

then restart. Pre-installed on comfy.icu.

Common issues

Red "missing" node in a shared workflow means the art-venture pack isn't installed - Manager's "Install Missing Custom Nodes" fixes it.

Now the LoRA problems that aren't this node's fault but will land on your desk. Nothing happens? Run the checklist: is the LoRA for the same architecture as your checkpoint (an SDXL LoRA does nothing on Flux, and even Pony/Illustrious/SDXL LoRAs cross over unreliably despite sharing code); is the weight so low it's inaudible; and does the LoRA need a trigger word in your prompt that you forgot. Black images or NaN errors when stacking? Add one LoRA at a time with a fixed seed, and if it still breaks, reorder with the larger LoRAs first. On a GGUF-quantized checkpoint LoRAs apply slower because each layer has to be dequantized, patched, and requantized - that's expected, not broken.

One honest note: if your real goal is stacking many LoRAs with per-entry toggles and trigger-word lookups, most people reach for rgthree's Power Lora Loader for that. AV_LoraLoader is a clean single-LoRA node you chain - perfect when you want a couple, with a mute switch, and nothing fancier.