Simple KSampler 🌡️

If you've ever run a diffusion workflow, this node is not going to surprise you - and that's the whole point. The Simple KSampler from TensorVizion/OmniNodes is a thin, branded wrapper around ComfyUI's own core KSampler. Same sampler and scheduler lists, same math, same behavior. The only real difference is a summary output that tells you exactly what ran.

Why does a wrapper node exist at all? The author's answer is architectural: so a complete workflow can be built without leaving the pack. This is OmniNodes' "one node every workflow needs" - paired with its Empty Latent Image, VAE Encode/Decode, and Simple SDXL Loader, you can go from checkpoint to image entirely inside the TensorVizion menu. You're not getting a better sampler; you're getting the standard one with a receipt.

The inputs

Every socket is a stock KSampler's, with tooltips that are genuinely worth reading:

• model - the denoising model. Wire it from a checkpoint, SDXL, or UNET loader.
• positive / negative - your conditioning, straight out of a CLIP Text Encode.
• seed - the noise's seed. The author's tooltip says it plainly: "the random seed used for creating the noise."
• steps - denoising steps. 20 is a fine starting point for SD-class models.
• cfg - classifier-free guidance. The tooltip is the classic warning: higher values stick closer to the prompt, but "too high values will negatively impact quality." If your image comes out fried, this is the first knob to lower - on a 2026 guidance-distilled model, the fix is usually CFG 1.
• sampler_name / scheduler - the pair everyone treats as one choice. dpmpp_2m + karras was the safe default for SD 1.5/SDXL years and still works there. If you're on a flow-matching model, that habit is wrong - take the model card's pair instead.
• denoise - at 1.0 it's a full generation from noise; drop it to 0.3–0.6 and you're doing img2img or a refinement pass from a meaningful latent.

What comes out

Two outputs. latent is the sampled latent, which you feed to a VAE Decode. summary is the interesting addition: a string reporting sampler, scheduler, steps, CFG, denoise, and seed - and it even classifies the run as "full generation" vs "partial denoise (img2img/refine)" based on your denoise value. Wire that into a text display or notes node if you want the graph itself to show what produced an image. It's a small thing, but it's the difference between "why did this render look like that?" and a self-documenting workflow.

Install

It ships with the OmniNodes pack, which installs like any custom node pack:

cd ComfyUI/custom_nodes
git clone https://github.com/TensorVizion/OmniNodes

Or use ComfyUI Manager (search OmniNodes), then restart. No extra dependencies - pure PyTorch. Note the category quirk: despite living in a "Sampling Nodes" folder, this node reports TensorVizion/Model Utilities as its category, so that's where you'll find it in the search menu. The pack's other sampling wrappers do the same; don't go looking under a dedicated Sampling submenu.

Troubleshooting

• Missing from the menu - restart ComfyUI fully, then check the terminal for the [OmniNodes] load log; ❌ Error importing shows a traceback you can chase.
• Results differ from a stock KSampler with the same settings - they shouldn't. This delegates straight to core; if you see a difference, check that you're feeding it the same values and the same model. A wrapper is a wrapper.
• Sampler choices look limited - the sampler_name enum carries 44 options (the brief samples only the first dozen); scroll it. It's core's full list, including the newer CFG++ variants.

Reach for this when you want one less thing to explain in a shared workflow. It's the same sampler you know, plus a paper trail.