KSamplerAdvanced [pipe] (inspire)

Two things are going on in this node's name, and both are the point. The [pipe] part means it takes a BASIC_PIPE - Impact Pack's bundle of model, CLIP, VAE, positive and negative conditioning on a single wire - instead of five separate connections. Cleaner graphs, less spaghetti. The Advanced part means you get manual step control: start step, end step, whether to add noise, whether to leave leftover noise. That's the combination you want for refiner passes and multi-stage sampling where you're handing a partially-denoised latent from one sampler to the next.

It also sits in Inspire's a1111_compat group, which is the pack's quiet superpower: it can generate noise the way Automatic1111 does (on the GPU), which is one of the main reasons ComfyUI and A1111 give different results from the same seed.

How it works

You feed it a BASIC_PIPE, a latent, and the usual sampler settings, and it runs an advanced sampling pass - respecting start_at_step / end_at_step so you can sample only a slice of the schedule. Because the pipe already carries the VAE, the node passes that straight through on its output, so you can wire a VAE Decode without re-routing the VAE separately.

The noise_mode is the interesting lever. GPU(=A1111) matches A1111's noise generation; CPU is ComfyUI's default behavior. There are also +internal_seed variants that mix in a separate internal seed. And batch_seed_mode controls how seeds spread across a batch - comfy applies noise to the whole batch at once (avoids accidental duplicate images), incremental bumps the seed per item like A1111 does, and the variation modes ramp variation strength across the batch.

The inputs and outputs

The knobs a beginner actually touches:

• basic_pipe (BASIC_PIPE) - your bundled model/clip/vae/conditioning, from a Make Basic Pipe node.
• noise_mode - set GPU(=A1111) if you're trying to reproduce an A1111 result; otherwise the default is fine.
• start_at_step / end_at_step and add_noise / return_with_leftover_noise - the step-range controls that make this the "advanced" sampler, for multi-pass and refiner setups.

The rest - steps, cfg, sampler_name, scheduler, noise_seed - are the standard sampling settings you already know. variation_seed / variation_strength let you keep a composition from one seed while nudging it slightly toward another (strength 0 = pure seed, 1.0 = pure variation seed). Optional inputs include noise_opt (feed in a NOISE_IMAGE, e.g. from a Seed Explorer node) and scheduler_func_opt. Outputs are a LATENT and the passed-through VAE.

How to install it

ComfyUI Manager: Install Custom Nodes → search ComfyUI Inspire Pack → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Inspire-Pack

One dependency matters: the BASIC_PIPE type comes from the Impact Pack (also Dr.Lt.Data's), so install that too or this node has nothing to accept. Both are one-click in Manager; restart after.

Common issues

If you can't find a source for basic_pipe, you haven't installed Impact Pack - that's where Make Basic Pipe and the pipe type live. Inspire's own Make Basic Pipe (Inspire) works too. Second, don't expect a plain KSampler and this node to produce byte-identical output; the whole reason to use it is noise_mode, and switching between GPU and CPU noise changes the result even at the same seed. And if your seeds behave oddly across a batch, check batch_seed_mode - incremental and comfy genuinely differ, and picking the wrong one is a common "why are my batch images not what I expected" cause.