FLUX.2 Klein Text/Ref Balance
One slider to trade prompt against reference
- model
- conditioning
- MODEL
- CONDITIONING
Every other reference-strength node in this pack scales one thing at a time - one reference, one channel, one mask. Text/Ref Balance is the odd one out: a single slider that trades prompt influence against reference influence as a pair, useful for the specific moment when you're not sure whether your edit isn't following the prompt because the reference is too dominant, or too weak.
How the slider actually behaves
This isn't a simple 0-to-1 crossfade the way it might sound. balance = 0.5 is neutral - at that point, text and references both sit at their normal scale of 1, untouched. Move below 0.5 and reference scale stays at 1 while text conditioning rises from 0 up to 1; move above 0.5 and text stays at 1 while reference scale falls from 1 down to 0. In other words, it's attenuation around a neutral midpoint in two separate directions, not an independent gain control that can push both streams above their normal strength simultaneously. You can weaken one side of the pair, but you can't boost both at once with this node.
That asymmetry matters for how you use it. If your edits feel like the reference is steamrolling the prompt, push the slider above 0.5 and the reference's scale backs off while text holds steady. If the model's ignoring your reference and just generating from the prompt, push below 0.5 and text backs off while the reference holds steady. Either direction, one side stays fixed at full strength while the other yields.
The inputs that matter
modelandconditioning- both required, both returned modified.balance- required, default0.5(neutral). Range 0–1, step 0.001 for fine control near the midpoint.debug(optional) - prints the resolved scale values, useful for confirming which direction you actually moved before trusting the visual result.
Outputs are a patched MODEL and the modified CONDITIONING - both need to go onward to your sampler.
Installing it
ComfyUI Manager: search ComfyUI-Flux2Klein-Enhancer, install, restart. Manual:
cd ComfyUI/custom_nodes
git clone https://github.com/capitan01R/ComfyUI-Flux2Klein-Enhancer.git
No extra dependencies.
Where people get burned
The mental model most people bring in is "0 = all reference, 1 = all text," like a crossfader - and that's not quite what's happening. Because one side always stays pinned at scale 1 while the other yields, this node can only ever pull one signal down relative to the other, never push either one up past its normal strength. If what you actually want is to boost the reference's pull beyond normal while keeping the prompt at full strength too, this node can't do that - you'd want Ref Latent Controller or Ref Latent Weight instead, which scale a reference independently rather than trading it off against text.
The fine step size (0.001) exists because the useful range tends to be narrow and close to 0.5 - small nudges near the midpoint do more than you'd expect, since you're already starting from both signals at full strength. Turn on debug the first time you use this so you can see the actual resolved values and build intuition for how far a given balance setting moves things before you start relying on it blind.
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| conditioning | CONDITIONING | — | |
| balance | FLOAT | 0.5000–1 | — |
| debugopt | BOOLEAN | false | — |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |
| CONDITIONING | CONDITIONING | — |