LCM Lora Loader

Small node, big reason it exists: this is what makes StreamDiffusion's "real-time" claim plausible instead of marketing copy. LCM-LoRA is a LoRA-shaped adapter distilled from Latent Consistency Models - drop it onto more or less any compatible checkpoint and that checkpoint can converge in 4-8 steps at CFG 1-2, instead of the usual 20-50 steps at CFG 5-9. That's not a small speedup, it's the whole reason real-time diffusion loops are possible on consumer hardware at all. This node's only job is loading that LoRA file and handing it to StreamDiffusion Create Stream, where it becomes a required input, not an optional extra.

It's worth knowing this isn't a general-purpose LoRA loader - it's typed specifically for this pack's StreamDiffusion pipeline, and functionally you're not going to get good results pairing it with anything that isn't already built around few-step generation. Skip this node and try to run StreamDiffusion on a full 20-50 step schedule and you've lost the entire point of the tool.

The inputs and outputs that matter

One input: lora_name, a dropdown of LoRA files you have available. One output: LCM_LORA, which plugs directly into StreamDiffusion Create Stream's lcm_lora input - nowhere else in the pack accepts this type.

How to install it

ComfyUI Manager → Install Custom Nodes → search ComfyUI-Diffusers → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Limitex/ComfyUI-Diffusers.git
cd ComfyUI-Diffusers
pip install -r requirements.txt

This node just loads a file you already have - the pack doesn't ship an LCM-LoRA itself, so you need to download one that matches your checkpoint's base architecture and drop it in your LoRA folder before this dropdown will show anything useful. Since you're building this specifically for StreamDiffusion, follow through on the rest of the README too: clone cumulo-autumn/StreamDiffusion inside the node's directory and run python -m streamdiffusion.tools.install-tensorrt if you want the accelerated path.

Common issues & troubleshooting

"Missing node type: LcmLoraLoader." Same fix as every other node in this pack: install ComfyUI-Diffusers via Manager or git clone, restart ComfyUI.

Empty dropdown, no LoRAs to pick. This node reads whatever's in your LoRA folder. If it's empty, you haven't downloaded an LCM-LoRA yet - go get one that matches your checkpoint's base (SD 1.5 LCM-LoRA on an SD 1.5 checkpoint, and so on; architecture mismatches don't error cleanly, they just produce bad output, the same trap as any other LoRA).

StreamDiffusion output looks blurry or generic at very few steps. This is the expected tradeoff, not a bug - the distillation knowledge base is blunt about it: LCM is best for rapid iteration and real-time preview, not final polished output. If you need higher fidelity, you're fighting the point of the technique; either accept the tradeoff for what real-time generation is for, or step back to a non-real-time workflow with a full step count for anything you actually want to keep.

Forgot to lower CFG. LCM-family adapters expect CFG in the 1-2 range - the StreamDiffusion sampler nodes in this pack default their own cfg to 1.2 for exactly this reason. Leaving CFG at a normal SD default (7-8) alongside an LCM-LoRA typically produces worse results than not using the LoRA at all, since the adapter was trained assuming low guidance.