Prompt Studio Sampler

Prompt Studio Sampler looks exactly like ComfyUI's stock KSampler, and that's not a coincidence. It is one. Open the source and the whole sample() method is a single call to ComfyUI's common_ksampler - the same helper the built-in KSampler runs. Same model, conditioning, and latent inputs, same LATENT output, same results. Swap it in and normal queues behave identically, bit for bit.

So why does it exist? One reason: Prompt Studio's XY(Z) plot builder needs a sampler it can safely drive, and it refuses to go poking random seed or cfg widgets elsewhere in your graph. That's a real design decision, not paranoia - an auto-detected "sampler" could silently rewrite some unrelated node that happens to share a widget name. The Pack owner's answer is an explicit contract: this node is the only thing the plot engine will touch for seed and sampling axes.

The mechanism is the marker, not the math

Here's the part that surprises people: the node adds zero sampling behavior. The actual denoising is delegated wholesale to common_ksampler from ComfyUI's own nodes module. The node itself is just a recognizable flag in the workflow that says "Prompt Studio may override these controls." The plot builder finds it by class name, exposes its seed, sampler, scheduler, steps, CFG, and denoise as axis options, and names the exact target node when you have more than one.

The KCPP prefix is a leftover from the pack's KoboldCpp heritage and it's a trap for beginners: this node calls no LLM, needs no server, and doesn't rewrite your prompt. It's pure diffusion sampling. The LLM business happens upstream in the pack's Prompt Slot / Prompt Amplify nodes.

When you'd actually reach for it

You'd use it inside a saved [PS] workflow when you want deterministic sampling controls exposed to Prompt Studio's XY(Z) plots - sweeping a sampler against a scheduler, walking CFG, or checking how denoise changes an img2img pass. If you don't run Prompt Studio, there is nothing here for you: installing this pack to get a nicer KSampler is wasted effort, because a stock KSampler in the same spot samples identically.

The inputs and output

They mirror the KSampler exactly, so there's no new vocabulary to learn:

• model, positive, negative, latent_image - wired the same way you wire any sampler.
• seed - with the standard control_after_generate randomize/increment behavior.
• steps (default 20), cfg (default 8.0), denoise (default 1.0).
• sampler_name and scheduler - populated from your ComfyUI build's lists (the sampler dropdown includes the CFG++ variants your install supports).

The single output is samples (LATENT), which feeds your VAEDecode just like the built-in node.

Installing it

This is just the ComfyUI_PromptStudio pack by tiko13. Easiest route: ComfyUI Manager → search ComfyUI_PromptStudio → Install → restart. Or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/tiko13/ComfyUI_PromptStudio
# restart ComfyUI

The pack ships no Python dependencies beyond ComfyUI itself - the pyproject declares none - so install is painless. The other features (prompt rewriting, the chat studio) need a local LLM at Ollama http://localhost:11434 by default, or a separately configured KoboldCpp/Llama.cpp server. This node needs none of that.

Where people get burned

• "I swapped it in and nothing changed." Correct - outside a Prompt Studio plot it's a passthrough. That's the feature.
• The workflow isn't listed in Prompt Studio. Saved workflows must start with [PS], contain a Prompt Slot or Amplify node, and have exactly one image output. A rejected one shows as cached - hover it for the validation reason, fix, and refresh.
• Multiple Prompt Studio Samplers in one graph. The plot axis builder names the exact target node, so pick the right one instead of assuming "the sampler."
• After any update, restart ComfyUI - the README's standing rule for this pack. If prompt creation works but no image appears, queue the workflow manually in ComfyUI and fix the graph there first.