🧱 Lora Loader - LD

The stock LoRA loader gives you one strength slider and assumes the file is all one thing. LTX-2 and newer audio–video models broke that assumption: since LTX-2 bet the family on synchronized sound, a LoRA trained on those models can carry video weights and audio weights in the same .safetensors. One slider means whichever half you didn't mean to boost comes along for the ride - a LoRA tuned to change a voice's character quietly rewrites the whole picture at the same strength. Lora Loader - LD exists because those two halves deserve separate faders. It is the pack's take on a LoRA controller, ported from the author's earlier PromptForgeLD pack and registered under LD / PromptMaster in PromptMasterLD.

How it works

Mechanically it's boring in the good way. Feed it your model and clip, hand it a stack_data list of up to ten LoRA slots, and it walks the stack in order, chaining each result into the next - model out of slot one goes into slot two, and so on down.

The trick is inside each slot. The loader opens the LoRA file and splits the weight dictionary by key name: anything with "audio" in the key goes to the audio pile, everything else is the video pile. Then it applies STR × V to the video weights and STR × A to the audio weights, calling ComfyUI's own load_lora_for_models twice - once per pile. So you can run a voice-character LoRA at A× 1.2 while leaving the picture at V× 0.6, or push the visual style hard and mute the audio side entirely by zeroing A.

You don't hand-write stack_data - it's a JSON string the node's own DOM widget builds for you. Ten rows, each with a searchable dropdown of every LoRA in your models/loras folder, an on/off toggle, and STR / V× / A× controls rendered as "cup of water" level boxes: hover and scroll to raise or lower, click to type a number. It behaves like the PromptForge loader it was ported from, which is the compliment it means.

Inputs and outputs that matter

Three inputs, two outputs - this is a small surface.

• model and clip - wire straight from your checkpoint or base-model loader. Outputs are model and clip again, so the node slots into the middle of a model rail like any other loader.
• stack_data - the slot list, driven by the on-node widget. Leave it alone unless you know what the JSON means.

Wire the output pair into a sampler (or another loader) and you're done. For an LTX audio-video workflow this replaces the "one LoRA at a time, then convert-and-merge" dance.

Installing it

It ships inside PromptMasterLD, so you get the whole pack whether you want it or not:

cd ComfyUI/custom_nodes
git clone https://github.com/Lora-Daddy/PromptMasterLD
# restart ComfyUI

No model files to download. The pack's only hard pip dependency is opencv-python (for its video-reference nodes, not this one); ComfyUI already ships the torch stack the loader touches. LoRAs themselves go in the standard ComfyUI/models/loras/ folder.

Common issues

• LoRA doesn't appear in the dropdown. It isn't in models/loras. Drop the file there and refresh; nothing scans subfolders you didn't create it in.
• Nothing visibly changes and the console says [LoraLoaderLD] LoRA not found. Missing files are a graceful pass-through, not a crash - which is friendly until you're debugging why a render ignored your style, so watch the console, not just the image.
• The audio half seems deaf. Some LoRAs are video-only - the file simply has no audio keys, so A× does nothing and there's nothing to split. That's expected, not a bug.
• Wrong render but the right look? If a LoRA is pulling in voice changes you didn't want, drop A× toward 0 before you throw the LoRA away. That split is the entire reason this node exists.