KSampler [pipe] (inspire)

This is the everyday version of Inspire's sampler, and it's genuinely nice to work with. Instead of running four or five wires into a KSampler - model, positive, negative, latent, and reaching for the VAE separately later - you hand it a single BASIC_PIPE that already bundles the model, CLIP, VAE and both conditionings. Your graph gets a lot tidier. Then it samples like a normal KSampler, with a denoise control and everything you'd expect.

The bonus, and the reason it lives in Inspire's a1111_compat group, is that it can make noise the A1111 way. ComfyUI generates its initial noise on the CPU; Automatic1111 does it on the GPU. That difference alone is one of the three big reasons a seed that looks great in A1111 comes out different in ComfyUI - this node lets you close that gap.

How it works

Feed it a BASIC_PIPE and a latent, set your sampler parameters, and it runs a standard denoise pass. Because the pipe carries the VAE, the node passes it straight through on the output - so you can go straight into VAE Decode without re-wiring the VAE. Unlike the Advanced sibling, this one uses a simple denoise slider rather than start/end step ranges, which is what you want for ordinary txt2img and img2img.

The noise_mode picks how noise is generated: GPU(=A1111) to match A1111, CPU for ComfyUI's default, plus +internal_seed variants. batch_seed_mode decides how seeds vary across a batch - comfy noises the whole batch together (dodges duplicate images), incremental steps the seed per item like A1111, and the variation modes ramp variation across the batch.

The inputs and outputs

The ones a beginner actually sets:

• basic_pipe (BASIC_PIPE) - your bundled model/clip/vae/conditioning.
• denoise - 1.0 for txt2img, lower (say 0.4–0.6) for img2img or a second pass.
• noise_mode - GPU(=A1111) when reproducing an A1111 render; default otherwise.

The rest - seed, steps, cfg, sampler_name, scheduler - are the familiar sampling settings. variation_seed / variation_strength hold a composition while nudging it toward another seed (0 = pure seed, 1.0 = pure variation seed). Optional scheduler_func_opt lets you pass a custom scheduler. 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

The BASIC_PIPE type comes from the Impact Pack (also Dr.Lt.Data's), so you'll want that installed too - otherwise nothing produces a pipe to feed this. Both install cleanly through Manager; restart after.

Common issues

Nowhere to plug in basic_pipe? Install Impact Pack, or use Inspire's own Make Basic Pipe (Inspire) - the pipe has to come from somewhere. Second, don't expect identical output to a stock KSampler at the same seed; the entire value here is noise_mode, and GPU-vs-CPU noise changes the image. Reproducing A1111 exactly also needs the other two pieces the pack's README calls out: prefix your embeddings with embedding:, and use BlenderNeko's Advanced CLIP Text Encode so the prompt weighting matches - noise mode alone gets you most of the way, not all of it.