LCMLoader

Every workflow in this pack starts here. LCMLoader builds the whole latent-consistency pipeline in memory and hands it to the generation nodes as a pipe - the thing that looks like an opaque "class" output. If this node errors, nothing else in the pack runs, so it's worth understanding what it's doing before you fight it.

What it actually loads

This is not a normal ComfyUI loader. It doesn't read a .safetensors through ComfyUI's checkpoint system. Instead it goes straight to HuggingFace diffusers and assembles the LCM_Dreamshaper_v7 model piece by piece - VAE, text encoder, tokenizer, UNet - then wraps them in a custom LatentConsistencyModelPipeline_inpaint. That's why the pack ships diffusers in its requirements and why the model has to sit in a special place, not in your checkpoints folder.

The model is a distilled one: LCM (latent consistency models) collapses the 20–50 step denoising trajectory into a handful of big jumps, so this pipeline is built for 4–8 steps, not 30. The KB's distillation essay has the honest verdict: it's for speed and iteration, not your final render.

The three inputs

• device - GPU or CPU. GPU is the intended path; it enables xformers memory-efficient attention and sequential CPU offload. CPU exists and mostly works, and it's actually the required choice for prompt weighting (see below).
• model_path - a string path. Leave it blank and it defaults to ComfyUI/models/diffusers/LCM_Dreamshaper_v7. You almost always leave it blank.
• tomesd_value - token merging, default 0.6. Higher merges more and speeds up inference at some quality cost. The defaults are fine; don't touch it until you're chasing speed.

Installing the model (this is where people get stuck)

The single most common failure for this whole pack, straight from the pack's own Reddit thread, is the loader dying with "We couldn't connect to 'huggingface.co' to load this model". That message means the diffusers folder isn't where the loader expects it. Fix it by hand-downloading the model once:

# the folder name must be exactly LCM_Dreamshaper_v7
mkdir -p ComfyUI/models/diffusers/LCM_Dreamshaper_v7
# grab the files from https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7/tree/main
# into that folder (vae/, text_encoder/, tokenizer/, unet/, feature_extractor/, model_index.json)

Then install the pack itself:

cd ComfyUI/custom_nodes
git clone https://github.com/taabata/LCM_Inpaint-Outpaint_Comfy
cd LCM_Inpaint-Outpaint_Comfy
pip install -r requirements.txt   # diffusers, tomesd, compel, icecream, Flask

Restart ComfyUI. Or skip all that and use ComfyUI Manager - search for LCM_Inpaint-Outpaint_Comfy. Manager won't download the model either, though.

Gotchas worth knowing

Windows path issues are real. The author's README warns that paths are formatted for Linux and errors happen on Windows - you'll see backslash/forward-slash bugs in the loader's default path logic. If you hit one on Windows, pass the full model_path string manually.

This is 2023-era code, and it shows its age: it loads the model into VRAM independently of ComfyUI's own cache, so you can end up with two copies of essentially the same weights in memory. For modern LCM work most people use ComfyUI's native LCM sampler with a normal checkpoint. This pack is the older, diffusers-based route - still fun, still fast, but not the tool most people reach for today. Use it because you want its inpaint/outpaint pipeline, not because it's the only way to run LCM.

The output is the pipe class connection - wire it into any LCMGenerate* node (the main one is LCMGenerate), and remember that one pipe can't be shared across two generation nodes simultaneously without reloading.