FluxGuidance (Mango)

If you've ever run a Flux workflow, you already know this node - FluxGuidance (Mango) is the same CFG-less guidance trick ComfyUI ships natively, with one extra wire bolted on. It takes your conditioning, stamps a guidance value into it, and hands that number to the sampler separately as a FLOAT output. In plain terms: it's the guidance node you already use, except the value survives long enough to land in your saved metadata.

Why does that matter? Flux Dev and Schnell are guidance-distilled models. There's no CFG and no negative prompt in the usual sense - you steer them with a single "guidance" number instead, and if you record the wrong thing in your image metadata, every Civitai recipe you post reads wrong. This node closes that gap: the sampler's metadata can point back at the exact guidance you actually ran, not a default the node never saw.

How it works

Under the hood it's a one-liner. The node calls conditioning_set_values with {"guidance": guidance}, which is precisely what ComfyUI's built-in FluxGuidance node does. The only difference is that it also returns the number you fed it (flux_guidance) as a FLOAT, ready to wire into a Mango sampler so it can be recorded in metadata. The mechanism is boring, and that's a compliment.

The inputs that matter

• conditioning - the CONDITIONING output of your CLIP Text Encode node. This is what gets the guidance value stamped into it.
• guidance - FLOAT, default 3.5, range 0–100. This is the one you'll actually fiddle with.

About that default: 3.5 is what every tool defaults to and it is genuinely controversial. Plenty of people argue it's too high and produces the "Fluxmaxxed" over-stylized face look, and a lot of the community lands at 1.8–2.0 for cleaner results - with the caveat that LoRA-heavy workflows sometimes do better pushed to 4.0–4.5. Treat 3.5 as a starting point, not a law. There's no need to touch the step size; 0.1 is fine.

The node has two outputs: conditioned (CONDITIONING), which feeds your Flux sampler, and flux_guidance (FLOAT), which you can leave dangling or wire into FluxSampler (Mango) to make your metadata honest.

Installing it

FluxGuidance (Mango) ships in the Mango Node Pack, so you get it the same way you get the rest:

cd ComfyUI/custom_nodes
git clone https://github.com/mang01010/MangoNodePack

Then restart ComfyUI. Or just search "Mango Node Pack" in ComfyUI Manager and click Install. The README lists pip install safetensors as the only extra dependency, though you'll almost certainly already have it in a stock install.

Gotchas

Because the guidance value rides along inside the conditioning tensor, the FLOAT output is your only easy way to read it back - don't forget to wire it into the sampler if you care about metadata. And one honest caveat about this pack in general: it's a small community project, so the nodes see less testing than the built-ins. If you just want to generate an image and never look at metadata, the native FluxGuidance node is a fine choice. Reach for this one when you want the paper trail.