CFE Flux In Pipe

CFE Flux In Pipe doesn't generate anything. It doesn't sample, it doesn't encode, it doesn't even convert. What it does is take six separate wires - model, clip, VAE, conditioning, sampler, scheduler - and braid them into one, so the middle of your Flux graph looks like a subway map instead of a plate of spaghetti.

That's the whole pitch of the pack. The author's README is one paragraph and it says the nodes exist "just for my own workflow" and are "a way of minimizing the noodles." This is the noodle-minimizer in its purest form: a bus collector at the start of a pipe line.

How it works

Under the hood it's embarrassingly simple - a tuple packer. All six inputs get put into a Python tuple in order:

pipe = (model, clip, vae, cond, sampler, scheduler)

That tuple is the FLUX_PIPE type, and one wire carries it across the canvas. Everything downstream that accepts a pipe (this pack's CFE FLUX Sampler (Pipe) and CFE Flux Out Pipe, among others) knows the order and unpacks it back out. No magic, no state, no caching - it just bundles.

One quirk worth knowing before you wire it: the pack treats sampler and scheduler asymmetrically. sampler is a real SAMPLER object, while scheduler is a plain string - the name of a scheduler, not an object. The author changed it to a string deliberately because it played nicer with their pipeline. The practical consequence is that you can't drop a stock scheduler dropdown straight onto the port; you need a node that emits the scheduler name as text. The pack's own CFE Sigma Sampler Strings exists precisely to turn a dropdown into the (SAMPLER object, scheduler string) pair this node wants.

The inputs that matter

• model - your Flux checkpoint. Whether it's full fp16 or a GGUF/fp8 quant, it goes here.
• clip - the T5 text encoder for Flux.
• cond - conditioning that's already been encoded. The pack's CFE FLUX Guidance node or a stock FluxGuidance + CLIPTextEncode chain both produce it.
• sampler / scheduler - the algorithm and noise schedule pair. They travel on the bus so the sampler end doesn't have to ask.
• vae - the VAE. Ignore its tooltip ("The latent image to denoise") - that's a copy-paste error from the model field, a detail the author clearly didn't care enough about to fix. It's just the VAE.

The single output, pipe, wires straight into CFE FLUX Sampler (Pipe) or CFE Flux Out Pipe.

Install

The README has no install section - the author cheerfully tells beginners it's "a sad place to be" not to know how. So, the standard route: ComfyUI Manager → search CFE_comfyui → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/CpreForEver/CFE_comfyui

Then restart ComfyUI. There are zero extra dependencies - no requirements.txt, no model downloads, no weights. The whole pack only imports things that ship with ComfyUI itself.

Where people get burned

Honestly, the main trap is expectations. This is a one-person playground pack, not a maintained library - no community footprint to speak of, no guarantees, and issues may sit unanswered. If you want a battle-tested pipe system, the big packs (rgthree, Efficiency Nodes) are the safe bet. But if you like the look of this one's minimal bus and you're already Flux-only, single-conditioning, no-negatives (which this pack assumes by design), it's genuinely unobtrusive. Garbage in, garbage out applies at bundle time too - the pipe doesn't validate what you put on it, it just carries it.