Nodes/ComfyUI-Nunchaku-Klein/Nunchaku FLUX.2 Klein Text/Ref Balance
ComfyUI Node

Nunchaku FLUX.2 Klein Text/Ref Balance

The text-vs-reference balance dial Klein needed

By tom-m-2020·Created 2 months ago·Updated 17 days ago· 2
Nunchaku FLUX.2 Klein Text/Ref Balance
  • model
  • conditioning
  • MODEL
  • CONDITIONING
balance0.500
debugfalse

A Klein edit is a tug-of-war between two instructions: what the reference image says and what your prompt says. The NunchakuKleinTextRefBalance gives you one slider for who wins. At balance 0.5 it's a fair fight; push it toward 1.0 and the references dominate, toward 0 and the text takes over. If you've ever had a reference image railroad your prompt ("I asked for a red shirt, why is it still blue?"), this is the node that fixes it.

Two things to know up front. First, the README is refreshingly honest about it: this is experimental and "not perfect," and it recommends balance in the 0.50-1.00 range - the low end behaves less predictably. Second, like the compat Ref Latent Weight, this is a prediction-space substitute rather than true attention K/V balancing. The Direct K/V version does the real thing but needs the extended Nunchaku fork. This one runs on stock Vitoom Nunchaku.

Inputs and wiring

model (from the pack's Model Loader), conditioning (with your reference latents attached via Multi Reference Latent), and balance (0-1, step 0.005). It outputs two things: a MODEL and a CONDITIONING - both flow onward to the sampler, since the balance touches both the model branch and the conditioning stream.

It cannot coexist with Ref Latent Weight on the same model branch - the source raises an explicit error. They're two different ways of steering the same thing, and the pack enforces "pick one." The natural pairing is with Mask Ref Controller, which constrains where references act while this node controls how hard.

Installing

The one pack install, repeated for the tenth time because it genuinely is the same for all of these:

cd ComfyUI/custom_nodes
git clone https://github.com/tom-m-2020/ComfyUI-Nunchaku-Klein

Vitoom Nunchaku wheel from huggingface.co/tonera/vitoom-nunchaku, tonera's Klein checkpoint in diffusion_models/, restart.

Common issues

  • Image sticks to the reference no matter what - you're below the sane range. Keep balance ≥ 0.5 per the README; below that the substitute math gets flaky.
  • "Cannot be stacked" error with Ref Latent Weight - remove one; they're mutually exclusive on a branch by design.
  • Warm cast on edits - balance isn't a color tool. Stack a Color Anchor for that; different problem.
CategoryNunchaku/FLUX.2 Klein/Enhancer/Reference

Inputs (4)

NameTypeDefaultDescription
modelMODEL
conditioningCONDITIONING
balanceFLOAT0.5000–1
debugoptBOOLEANfalse

Outputs (2)

NameTypeDescription
MODELMODEL
CONDITIONINGCONDITIONING