Diffusers Sampler

This is where a diffusers-pipeline workflow actually produces an image - the equivalent of ComfyUI's own KSampler, just running through HuggingFace's diffusers library underneath instead of ComfyUI's native inference code. If you've built the chain - Pipeline Loader → (optional Scheduler/VAE swap) → Model Makeup → Clip Text Encode - this is the last stop. Width, height, steps, CFG, seed, hit queue, get an image out.

There's nothing exotic about the mechanism itself: standard txt2img denoising, guided by your encoded positive and negative embeddings, walking down whatever scheduler you configured back in Diffusers Scheduler Loader (or the checkpoint's default if you didn't touch it). The reason this node exists as a separate thing from KSampler isn't that it does anything different - it's that it's built to run against a MAKED_PIPELINE object, and that's the only type this whole side of the pack understands.

The inputs and outputs that matter

Required:

• maked_pipeline - from Diffusers Model Makeup.
• positive_embeds / negative_embeds - from Diffusers Clip Text Encode (the EMBEDS outputs specifically, not the plain-text passthrough).
• width / height - default 512×512, same story as any SD-family model: stay near what the checkpoint was trained at or you'll get the usual duplicated-limb, tiling artifacts that come from generating well above native resolution.
• steps - default 20.
• cfg - default 8, the standard SD 1.5/SDXL range (roughly 5–9 is the community sweet spot; push much past 12 and you start seeing oversaturation and burned colors).
• seed - fixed or randomized per run.

One output: IMAGE.

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

No extra weights beyond the checkpoint you already loaded in Diffusers Pipeline Loader.

Common issues & troubleshooting

"Missing node type: DiffusersSampler." The pack isn't installed, or ComfyUI hasn't been restarted since you installed it. Install via Manager or git clone, then restart - this is the fix for every missing-node report on this pack's nodes.

Grey, washed-out output. Same root cause as it is everywhere in the ecosystem: a missing or mismatched VAE. If you routed a VAE override through Diffusers VAE Loader, confirm it actually matches the checkpoint's architecture - an SD 1.5 VAE on an SDXL-family checkpoint gets you a poor decode, not an error, so it's easy to miss.

Weird artifacts at unusual aspect ratios. Width/height defaults to 512×512 for a reason - that's SD 1.5 territory. Push an SDXL-based checkpoint through this node at the same defaults and you'll likely see the classic "trained-resolution mismatch" symptoms: duplicated anatomy, tiling, stretched limbs. Match width/height to whatever the underlying checkpoint actually expects.

Image looks destroyed / noisy at the end of generation. Classic too-high-CFG-with-too-few-steps combination. If cranking CFG up didn't get you the prompt adherence you wanted, add steps rather than pushing CFG further - CFG in the high teens and beyond tends to burn out an image regardless of which sampling stack is running underneath.

It won't connect to anything from your native ComfyUI graph. MAKED_PIPELINE, EMBEDS, and this node's IMAGE output are all fine to use downstream normally, but the inputs here only accept this pack's own upstream types - you can't feed it a native MODEL or CONDITIONING. If you're not otherwise using the Diffusers node family, the built-in KSampler is almost certainly what you want instead.