Diffusers Scheduler Loader

Quick terminology trap first: this isn't ComfyUI's sampler/scheduler split (the KSampler dropdown pair you're probably used to, euler + karras and so on). This node is picking a diffusers library scheduler class - DDIM, DPMSolverMultistep, UniPCMultistep, that family - because everything in this pack routes through HuggingFace's diffusers Python package instead of ComfyUI's own sampling code. Same underlying idea - an algorithm that decides how a pipeline steps from noise to image - different naming convention, different implementation, and it only plugs into other nodes in this pack.

You take it in a pipeline (from Diffusers Pipeline Loader) and swap in whichever scheduler you want to use instead of the checkpoint's default, then it hands that scheduler on to Diffusers Model Makeup for final assembly.

The inputs and outputs that matter

• pipeline (required) - the loaded diffusers pipeline this scheduler will attach to. The node needs it to configure the scheduler correctly against that pipeline's config.
• scheduler_name (required, dropdown) - one of eleven diffusers scheduler classes: DDIM, DDPM, DEISMultistep, DPMSolverMultistep, DPMSolverSinglestep, EulerAncestralDiscrete, EulerDiscrete, HeunDiscrete, KDPM2AncestralDiscrete, KDPM2Discrete, UniPCMultistep.

Output is a single SCHEDULER, which feeds into Diffusers Model Makeup alongside the pipeline and (optionally) an overridden VAE.

If you're not sure which to pick: DPMSolverMultistep and EulerDiscrete are the most commonly reached-for general-purpose choices in the diffusers ecosystem - DDIM is the oldest and slowest to converge but the most "reference" implementation, and UniPCMultistep tends to do well at low step counts. None of this is ComfyUI folklore territory (Karras vs. non-Karras, ancestral vs. converging) carrying over automatically - these are diffusers' own scheduler implementations, and their step/quality tradeoffs are documented on HuggingFace's own scheduler pages, not in ComfyUI community threads.

How to install it

ComfyUI Manager → Install Custom Nodes → search ComfyUI-Diffusers → install → restart. Or manually:

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

That covers this node. No model weights involved - schedulers are just algorithms, nothing to download.

Common issues & troubleshooting

"Missing node type: DiffusersSchedulerLoader." The usual story for this pack - you loaded someone else's workflow, don't have ComfyUI-Diffusers installed, and ComfyUI reports the bare class name with no hint where it's from. Users have hit this exact pattern with sibling nodes in the pack and had to search the web to trace it back to Limitex's repo. Install the pack via Manager and restart; that's the fix every time.

Scheduler swap does nothing visible. Make sure the SCHEDULER output actually feeds into Diffusers Model Makeup, and that Model Makeup's MAKED_PIPELINE output (not the original pipeline) is what you're sampling from downstream. It's an easy wire to drop when you're first wiring this pack together, since the graph has more hops than the native checkpoint → KSampler path you're probably used to.

Results look worse than the checkpoint's default. Not every scheduler suits every model or step count - some need more steps to converge cleanly (DDIM), some are tuned for fewer (UniPCMultistep). If a swap makes output noticeably worse, that's often just a step-count mismatch rather than a bug; try adjusting steps in Diffusers Sampler before assuming something's broken.

It won't connect to anything outside this pack. SCHEDULER is a diffusers-pack type, not a ComfyUI native one - it stays inside the Diffusers/StreamDiffusion node family. If your workflow is otherwise native ComfyUI, you almost certainly want the built-in KSampler's scheduler dropdown instead of this node.