ComfyUI-Nunchaku-Klein

ComfyUI custom node only for tonera's Nunchaku Klein model using the community-maintained Nunchaku backend.

Additional features such as Enhancers (not only ported but compatibility versions with substitute algorithms + Direct K/V versions requiring a farther extended nunchaku backend) are being developed.

Work in progress.

Check the GH repo for more detail.

News

  • 2026-08-11 v0.7.0
    • Ref Latent Weight (Direct K/V)
      • NOTE: NEEDS NEW NUNCHAKU BACKEND. To use this node you'll need to reinstall nunchaku with our "fork of the community-maintained Nunchaku" that means, Vitoom Nunchaku which is an extended version of nunchaku by tonera has been farther extended by me, which gives you full control over Flux.2 Klein.
  • 2026-08-09
    • v0.6.0
      • Text/Ref Balance (not perfect, experimental. Use 0.50-1.00 in balance.)
    • v0.5.0
      • Ref Latent Weight
  • 2026-08-08 v0.4.1
    • Multi Reference Latent
    • Mask Ref Controller
    • Color Anchor
  • 2026-08-06 v0.3.0
    • Sectioned Encoder
    • Detail Controller
  • 2026-08-05 v0.2.0
    • Klein Enhancer
  • 2026-08-04
    • Foundation and basic functionalities for 9B have been mostly complete.
    • LoRA support
    • Multi-reference editing support
    • First experimental implementation of Klein Text Enhancer node.
      • NOTE: Our "Enhancer" nodes do not work exactly the same as the original Flux2Klein-Enhancer, but the behavior is close. Still experimental.
    • 4B not supported yet.

Observed: repeated Nunchaku FLUX.2 Klein executions can produce materially different outputs despite identical workflow inputs and seed. This also reproduces in plain T2I without references, Ref Latent Weight, or LoRA, so it is not specific to those target features.

Todos:

  • 9B basic support
  • LoRA support
  • Reference Edit support
  • Klein Enhancer (partially)
  • Test Differential Diffusion / if not: SpotEdit
  • Test Normalized Attention Guidance
  • 4B support

Installation

Basic Usage

  1. First you need to install tonera's fork of nunchaku (Vitoom Nunchaku). Choose the pre-built wheel matching your setup from their repository.
  2. Clone this repository.
  3. Download tonera's Nunchaku Klein model from their repository if you don't have one.

Direct K/V

If you want to use Ref Latent Weight (Direct K/V), Text/Ref Balance (Direct K/V) or Identity Feature Transfer:

Either:

  • Install my "fork of the fork" of nunchaku, or from HF repo.
  • Use "repack" script in the repo to your chosen pre-built wheel (if the version exactly match.)
    • If you use the repack script, they MUST match the exact variant of pre-built Vitoom Nunchaku wheel as the source e.g.:
      • If you have nunchaku-1.3.0.dev20260629+cu13.0torch2.11-cp313-cp313-win_amd64, you must use the script that targets nunchaku-1.3.0.dev20260629+cu13.0torch2.11-cp313-cp313-win_amd64.

Currently supported wheels

Only a limited number of versions are supported now.

  • nunchaku-1.3.0.dev20260629+cu13.0torch2.11-cp313-cp313-win_amd64

(Basic usage supports all the other versions, too.)

Environment

Tested with:

  • ComfyUI >=0.29. Older versions may not work.
  • Python 3.13
  • Torch 2.11
  • CUDA 13.0

Related Projects

This project involves/is related to several third-party projects

LICENSE

Custom Node

GPL-3.0-or-later

Model Weights

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support