Load Loras From Prompt

If you spent any time in Automatic1111 before moving to ComfyUI, you know the drill: type <lora:add-detail:0.8> right into your prompt box and the extension quietly loads that LoRA at 0.8 strength before generation. Load Loras From Prompt brings exactly that habit into ComfyUI. Instead of dragging in a LoRA loader node and hunting through a dropdown for the right file, you write the tag straight into your prompt text, and this node reads it, applies it, and strips it back out before the rest of your prompt goes anywhere near CLIP encoding.

It's part of the Comfy-Photoshop-SD pack, which exists to bridge ComfyUI with Abdullah Alfaraj's Auto-Photoshop-SD plugin - a Photoshop extension that fires ComfyUI workflows over the API. That context explains the design choice here: the plugin's prompt box is just a text field. Building a separate LoRA-picker UI inside Photoshop would be real work; parsing tags out of the same text field the artist is already typing into is nearly free. So instead of a graph node with a dropdown, you get a node that reads your prompt the way A1111 always did.

Inputs and outputs

Three required inputs: model and clip (the pair straight off your checkpoint loader), and prompt - a multiline string where you write your actual prompt, LoRA tags and all. Three outputs come back: MODEL and CLIP with the LoRAs baked in, ready to hand to your sampler and text encoders, and a STRING - your prompt with the <lora:...> tags removed, clean and ready to drop straight into a CLIPTextEncode node. You don't wire the raw prompt into CLIPTextEncode; you wire this node's STRING output.

Mechanically, the workflow is: CheckpointLoaderSimple → this node's model/clip inputs, your prompt text → this node's prompt input, and this node's MODEL/CLIP/STRING outputs go on to do exactly what the checkpoint's model and clip and your prompt string would have done anyway - just with the LoRAs already folded in along the way.

The honest take

If you're ComfyUI-native, you'll probably prefer something like rgthree's Power Lora Loader - a proper node with toggles, per-entry weight sliders, and a right-click that pulls trigger words off the file or CivitAI for you. That's the more discoverable, more forgiving way to stack LoRAs in a visual graph. This node earns its place specifically when your prompt text is coming from somewhere that isn't a ComfyUI node editor - an API caller, a saved template, or exactly the Photoshop plugin it was built for - where typing a tag into an existing text field is much less friction than wiring in a new node.

Installing it

Through ComfyUI Manager: search "Comfy-Photoshop-SD", install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/AbdullahAlfaraj/Comfy-Photoshop-SD

then restart ComfyUI. If you're planning to use any of this pack's ControlNet nodes too, the README also asks you to install Fannovel16's comfyui_controlnet_aux - not needed for this specific node, but worth doing in the same pass if you're setting the whole pack up.

Where people get burned

The LoRA name in your tag has to match the filename in ComfyUI/models/loras exactly, minus the extension - no fuzzy matching, no partial names. A typo doesn't give you a friendly warning banner in the graph; it either fails silently or throws in the console, so if a LoRA you tagged clearly isn't affecting the output, that's the first thing to check. Second, keep the weight a plain number (0.8, not 80% or anything cute) - the same convention A1111 trained everyone on, and this node follows it.