KSampler

The first time you spot this in the node browser it looks like a mistake - why does an LLM-agent pack ship its own copy of ComfyUI's core KSampler? The answer is right there in comfyui_LLM_party's own pitch: it wants to let people "easily integrate their LLM workflows into their existing image workflows" and build "one-stop LLM + TTS + ComfyUI" pipelines. Shipping a copy of the sampling node inside its own category means an agent-built graph never has to reach outside the pack's own node tree to finish the image half of the job.

Functionally, it's the same node you already know. Same required inputs, same behavior, same output. There's no hidden "party" trick here - this is a straight re-implementation of core sampling, not an enhanced version.

How it works

Exactly like stock KSampler: it takes a model, conditioning, a starting latent, and the usual set of sampling controls, runs the diffusion sampling loop, and returns a denoised latent.

The inputs and outputs that matter

• model (MODEL) - your loaded checkpoint.
• positive / negative (CONDITIONING) - the usual encoded prompt pair.
• latent_image (LATENT) - the starting latent (empty latent for txt2img, an encoded image for img2img-style denoise).
• seed, steps (default 20), cfg (default 8), denoise (default 1) - the standard numeric knobs, same defaults as core ComfyUI.
• sampler_name - a full 44-option list covering everything from euler through the exotic DPM variants.
• scheduler - the usual 9 options (simple, karras, sgm_uniform, and so on).

Output: LATENT - feed it into your VAE Decode, same as always.

Should you use this instead of the real KSampler?

Honestly, it doesn't matter much either way - there's no functional difference. The one real reason to reach for this one specifically: if you're building an entire pipeline inside this pack's node categories (an LLM node writing your prompt, a tool node fetching data, this sampler turning it into an image), keeping everything under one category tree is a little tidier to navigate than jumping back and forth between this pack's nodes and core ComfyUI's. If you're not doing that - if you're just running a normal image workflow with one LLM step bolted on - stick with the KSampler you already have wired up. Swapping it for this one buys you nothing.

How to install it

Search comfyui_LLM_party in ComfyUI Manager and install, then restart. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/heshengtao/comfyui_LLM_party.git

Run pip install -r requirements.txt from inside the pack's folder using ComfyUI's own Python, then restart. This node needs the same underlying PyTorch/diffusion stack any sampler needs - nothing specific to this pack beyond that, since it's a re-implementation of functionality ComfyUI core already ships.

Common issues & troubleshooting

Whatever debugging habits you already have for the stock KSampler apply here unchanged - this is the same sampling loop under a different node name. If your results look wrong (blown-out images, no denoise happening, garbage output), check the usual suspects first: cfg too high or low for your model family, denoise left at 1 when you meant a partial img2img pass, or a sampler_name/scheduler combination your checkpoint doesn't handle well. None of that troubleshooting changes because you're using this pack's copy instead of core's.