Stream_Lora_Loader

Stream_Lora_Loader sits between Stream_Model_Loader and Stream_Sampler, and it's the least optional node in the pack. It's the hop where you bolt a style LoRA onto the pipeline - on top of the LCM LoRA the model loader already baked in - so the sampler can actually render in the style you wanted at real-time-ish speed. Nothing about the node is complicated; the sneaky part is that it doubles as a message carrier for the whole chain.

Mechanically it's three diffusers calls. It loads your style LoRA with the adapter named after your trigger_word, sets the adapter weights to [1.0, lora_scale] (the LCM LoRA stays at full strength, your style LoRA at the scale you picked), then fuses the two LoRAs into the base weights. Because the adapter name is your trigger word, that word is load-bearing: the sampler later appends prompt + trigger_word + "style" to whatever you type. So when the README says "when you change a style LoRA, the trigger word needs to change too" - it's not lore, it's literally wired into the code. Get the word wrong and your prompt is just missing a token; the style barely shows.

The inputs you actually set:

• lora - your style LoRA, from models/loras. Match it to the base model: an SD1.5 style LoRA on an SDXL checkpoint does nothing useful, and a LoRA is a patch, not a model (if you need the distinction, the modidex concepts doc walks through it).
• lora_scale - 0.1 to 1.0, default 0.8. Style LoRAs get mushy at 1.0; the classic 0.6–0.8 sweet spot is right here.
• trigger_word - default "best quality", but you'll replace it with the trigger from the LoRA's card (or leave the card's word blank and just describe the style).

It consumes the model loader's pipe and info, and emits a pipe plus a rewritten info. The info string now carries three fields (model_type;trigger_word;checkpoint) because the sampler expects exactly three. That's the practical trap: this node is effectively required even if you don't want a style LoRA - the code raises "you need a style lora" if you leave lora on "none". And if you try to skip the node and wire Stream_Model_Loader straight to Stream_Sampler, the sampler's info.split(";") will crash on the missing field. So keep it in the chain; if you truly want no style LoRA, the "none" route isn't really an option, which is the pack being honest about what it's for.

Install and models are shared with the rest of the pack: Manager search "ComfyUI_Streamv2v_Plus" or git clone https://github.com/smthemex/ComfyUI_Streamv2v_Plus into custom_nodes, restart, pip install -r requirements.txt (expect to also add diffusers, transformers, torchvision, opencv-python - the author's own fix-it-as-you-go install). The known pain point is the PEFT error: "PEFT backend is required for this method" means update with pip install -U peft transformers.

Realistically this node is where most people's "why doesn't my style show" questions live, and the answer is almost always either the trigger word not matching the LoRA or the LoRA's base family not matching the checkpoint. Fix those two and the node quietly does its one job: style applied, chain passed forward, zero drama.