ComfyUI FDA — Fourier Domain Adaptation

A custom ComfyUI node that applies Fourier Domain Adaptation (FDA) between a target image and a reference image. FDA transfers the global color and style of the reference into the target by swapping low-frequency amplitude components in the frequency domain, while fully preserving the structural content of the target.

---

Screenshots

---

How It Works

1. Apply 2D FFT to each channel of both images.
2. Shift the zero-frequency component to the center of the spectrum.
3. Decompose each spectrum into amplitude and phase.
4. Replace the low-frequency amplitude region of the target with that of the reference. The region size is controlled by beta.
5. Reconstruct the complex spectrum and apply the inverse FFT.

The result inherits the structure (edges, shapes) of the target and the color/tone statistics of the reference.

---

Installation

Copy the ComfyUI_FDA folder into your ComfyUI custom_nodes directory and restart ComfyUI.

cp -r ComfyUI_FDA /path/to/ComfyUI/custom_nodes/

No additional dependencies are required beyond the standard ComfyUI environment (torch, numpy).

---

Node

Category: image/transform Display name: Fourier Domain Adaptation (FDA)

Inputs

Name	Type	Description
target_image	IMAGE	The image whose structure you want to preserve.
reference_image	IMAGE	The image whose color/style you want to transfer.
beta	FLOAT	Low-frequency window size as a fraction of image dimensions. Range: 0.001 – 0.5. Default: 0.01.

Output

Name	Type	Description
adapted_image	IMAGE	The target image adapted to the style of the reference image.

---

Parameter Guide: beta

beta controls how large the low-frequency window is relative to the image size.

beta	Effect
0.001 – 0.01	Subtle color/tone shift; fine details unchanged.
0.01 – 0.05	Moderate style transfer; noticeable but natural-looking.
0.05 – 0.2	Strong style transfer; global color palette heavily influenced.
0.2 – 0.5	Aggressive transfer; may introduce visible artifacts.

Start with the default (0.01) and increase gradually to taste.

---

Batch Behavior

• If the target batch size (B_t) and reference batch size (B_r) differ, reference frames are cycled with i % B_r.
• If the reference image has different spatial dimensions from the target, it is automatically bilinearly resized before processing.

---

Reference

Yang, Y., & Soatto, S. (2020). FDA: Fourier Domain Adaptation for Semantic Segmentation. CVPR 2020.