Load Diffuser Model

This is where the outpaint pipeline starts: it loads the SDXL unet that's going to do the actual denoising. It's the diffusers-pack equivalent of ComfyUI's own Load Diffusion Model node, and it exists as its own thing for one reason - the rest of this pack runs a diffusers pipeline internally rather than ComfyUI's native model stack, and diffusers wants its model config laid out differently than ComfyUI's .safetensors loader expects.

Don't confuse this with a full checkpoint loader. The README is explicit that this node "supports only unets" - if you want to load a whole checkpoint (unet + clip + vae bundled together), use ComfyUI's own Checkpoint Loader Simple node instead and just feed its CLIP/VAE outputs into the rest of your graph. This node only wants the unet file.

What you actually set: unet_name is a dropdown populated from whatever's sitting in your ComfyUI/models/diffusion_models folder - it'll be empty until you put something there. The pack's own README only vouches for one model here: RealVisXL_V5.0_Lightning, an SDXL Lightning build, downloaded as the diffusion_pytorch_model.fp16.safetensors unet file rather than a full checkpoint. model_type defaults to sdxl - leave it there; the pack's configs are hardcoded for SDXL only (the author explicitly gave up on getting Flux working), so it's really a formality rather than a real choice. device and dtype default to auto, which lets ComfyUI pick sensibly - there's rarely a reason to override these unless you're deliberately forcing fp16/bf16 for a VRAM-constrained card.

Outputs: model and scheduler configs (a SCHEDULER type, not something you'll see elsewhere in ComfyUI) both feed straight into the Diffusers Image Outpaint node's model and scheduler_configs inputs. You need both - this node hands them over as a pair, since the diffusers pipeline needs the scheduler config that matches the unet it just loaded.

Installing it: it comes bundled with the rest of the pack - git clone ComfyUI-DiffusersImageOutpaint into custom_nodes (or install via the Manager) and restart. If the Manager didn't already pull the Python dependencies, run pip install -r requirements.txt from inside the extension folder using your ComfyUI environment. The node itself needs no extra setup beyond having a model file to point at.

Where this trips people up: an empty dropdown almost always means you haven't put a unet file in diffusion_models yet - the folder isn't seeded with anything by default, and this extension does not auto-download models for you. If you load something other than RealVisXL_V5.0_Lightning and get a cryptic error downstream, remember the author has only actually tested that one model - since the model-type configs here are hardcoded to SDXL, an oddly-shaped SDXL unet (or, worse, anything Flux) is likely to fail rather than gracefully degrade. And if you hit a missing 'loaded_keys' error the moment generation starts, that's not this node's fault specifically - it's a known break in newer transformers/diffusers releases, fixed by pinning pip install transformers==4.45.0 --upgrade diffusers==0.32.2 --upgrade inside your ComfyUI venv.