LoraLoaderTextRandom

LoraLoaderTextRandom is LoraLoaderText with the "variety" dial turned all the way up. Same text-driven lora_name, same MODEL/CLIP outputs - but instead of fixed strengths, it rolls both the model strength and the CLIP strength from ranges you set, on every generation. Random LoRA, random intensity, every time.

That's a genuinely useful mode for exploration. Fixed strengths are how you get a consistent style from a character LoRA; randomized strengths are how you survey what a LoRA can do without sitting there hand-turning a slider across thirty generations. Set a wide band (0.5 to 1.0 on model strength) and batch - you'll see the sweet spot and the blowout range in one pass. People were asking for exactly this - "randomize lora values would be awesome" - in the threads that first surfaced this pack in 2023.

Inputs and outputs

• model (MODEL), clip (CLIP) - passthroughs, returned patched.
• lora_name (STRING) - free-text filename, so you can wire in a wildcard expansion and roll the which LoRA too.
• seed (INT) - accepted, but here's a source-level quirk: the code never actually calls random.seed(seed). The strength roll isn't seed-pinned, so treat seed as decorative on this node and don't chase reproducibility with it.
• strength_model_min / strength_model_max (default 0.5–1.0) and strength_clip_min / strength_clip_max (default 0.5–1.5) - the ranges. The node draws random.uniform(min, max) for each, and it sorts them defensively, so flipping the min and max doesn't error.

Outputs: patched MODEL and CLIP. Like its sibling, a missing or empty LoRA name silently passes the inputs through.

Install

Clone and restart; ComfyUI Manager finds it by searching "Lilly":

cd ComfyUI/custom_nodes
git clone https://github.com/lilly1987/ComfyUI_node_Lilly

Restart after cloning. No model downloads; rich and chardet auto-install on first load.

Caveats worth internalizing: randomizing CLIP strength wide (the default top is 1.5) is the fast road to oversaturated, style-melted output - keep the clip band tighter than the model band unless you enjoy surprises. And remember this pack is unmaintained; the author now points to Impact Pack. If you want maintained random-LoRA behavior, later packs built the same idea with proper seed handling, but for an old workflow already on Lilly, this node still does the job.