CFE FLUX Sampler

CFE FLUX Sampler is the heart of this pack: a purpose-built sampling node that does everything the stock KSampler does, except it handles Flux's guidance the way Flux actually wants - one conditioning, no negatives, no classifier-free-guidance double pass.

If that sounds like a niche way to save one wire, it is. But it's also the difference between understanding what your sampler is doing and just clicking a dropdown. And for Flux specifically, the "no negatives" thing isn't a stylistic choice - it's the architecture. Flow-matching models like Flux were trained with guidance handled internally; the community shorthand is that Flux wants a guidance value (3.5-ish for Dev, 1.0 for guidance-distilled variants) rather than a CFG scale with a negative prompt attached. This node encodes exactly that philosophy.

How it works

The sampler leans on the pack's _flux_sampler helper, which is basically a hand-rolled KSampler. The interesting parts:

• It calls fix_empty_latent_channels on your latent so an empty latent from an aspect-ratio or Empty Latent node just works.
• It writes the cfg value straight into your conditioning as the guidance field.
• It builds a CFGGuider with only the positive conditioning - no negative slot anywhere.

Then it runs: guider.sample(noise.generate_noise(latent), latent, sampler, sigmas, seed=noise.seed). The NOISE input is ComfyUI's stock RandomNoise node - that's where the seed and the noise pattern come from. Everything about how many steps and what schedule is already baked into the SIGMAS input, which is why this node has no steps field. You build your sigmas separately (this pack's CFE Sigma Sampler is the obvious source) and the sampler just consumes them.

Inputs and outputs that matter

The full required list: noise, model, vae, cond, sampler, sigmas, latent_image, and cfg.

The two you'll actually touch:

• cfg - default 1.0, range 0–100. This is your Flux guidance value. Here's where people get burned: 1.0 is the correct default for guidance-distilled models (Schnell, most step-distilled LoRA stacks) but too low for Flux Dev, which wants roughly 3.5. Leave this at 1.0 on a Dev checkpoint and you'll get flat, under-committed images and blame the node. It's not the node.
• sigmas - comes from CFE Sigma Sampler (or any SIGMAS producer). If you feed it garbage sigmas you get garbage output; the sampler trusts you.

Outputs are latent (LATENT) and vae (VAE passthrough). Wire latent into a VAE Decode and you're done.

Install

Zero dependencies - the pack only uses ComfyUI's own imports, no model downloads:

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

Restart, then find it in the CFE/flux menu. Or use ComfyUI Manager → search CFE_comfyui.

The honest take

This is the one node in the pack I'd actually reach for over the stock KSampler if I were already running the pack's pipe system. It's clean, it's explicit about the Flux guidance model, and the single-conditioning design matches how Flux is supposed to run. That said, stock ComfyUI's own FluxGuidance + KSampler combo does the same job with no extra install - you're paying for tidiness, not capability. And this being a one-person playground pack, don't expect updates or issue responses if something breaks. For a beginner the real lesson is in the defaults: on Flux Dev, set cfg to 3.5, and never touch the negative prompt, because this node doesn't have one and Flux doesn't want one.