ComfyUI DCT Nodes

A ComfyUI custom node pack for 2-D Discrete Cosine Transform (DCT) based frequency-domain filtering. Draw a mask directly on the DCT spectrum inside the node to zero out any frequency band, then reconstruct the filtered image with the inverse transform.

---

Nodes

1 · DCT 2D Transform

Converts a spatial-domain image into its 2-D DCT-II coefficient map.

Input	IMAGE
Outputs	dct_spectrum — raw DCT coefficients [B, H, W, C] (custom DCT_SPECTRUM type)
	spectrum_preview — log-magnitude visualisation normalised to [0, 1] (IMAGE)

The transform uses scipy.fft.dctn with norm='ortho' (orthonormal DCT-II). Low frequencies sit in the top-left corner; higher frequencies extend toward the bottom-right.

---

2 · DCT Spectrum Mask Editor

Applies a user-painted binary mask to the DCT coefficient tensor. Zeroed coefficients are removed from the spectrum before reconstruction.

Input	dct_spectrum (DCT_SPECTRUM)
Outputs	masked_spectrum — filtered DCT coefficients (DCT_SPECTRUM)
	mask_preview — greyscale mask image (IMAGE)

Built-in Paint Canvas

The node embeds an interactive paint canvas directly in the ComfyUI graph. After running the upstream DCT 2D Transform node, the spectrum preview is automatically shown as the canvas background.

Control	Action
Left-click / drag	Paint black → zero out those DCT coefficients
Right-click / drag	Paint white → restore those coefficients
Scroll wheel	Increase / decrease brush size (2 – 120 px)
Double-click	Clear the mask (fill white = keep all frequencies)

---

3 · Inverse DCT 2D

Reconstructs the spatial-domain image from (masked) DCT coefficients.

Input	dct_spectrum (DCT_SPECTRUM)
Output	reconstructed_image — recovered image clipped to [0, 1] (IMAGE)

Uses scipy.fft.idctn with norm='ortho'. Without any mask applied the round-trip error is at machine-epsilon level (< 4 × 10⁻¹⁶).

---

Recommended Workflow

[Load Image]
     │
     ▼
[DCT 2D Transform] ──► spectrum_preview (optional preview node)
     │
     │ dct_spectrum
     ▼
[DCT Spectrum Mask Editor]   ← paint mask on the spectrum canvas
     │
     │ masked_spectrum
     ▼
[Inverse DCT 2D]
     │
     ▼
[Preview Image / Save Image]

Frequency Filtering Examples

Painted area	Filter type
Top-left region (low freq)	High-pass — sharpens edges, removes flat background
Outer region (high freq)	Low-pass — blurs image, removes noise and fine detail
Custom shape	Arbitrary frequency-domain filter

---

Installation

1. Clone or copy this folder into ComfyUI/custom_nodes/:

   cd ComfyUI/custom_nodes
   git clone https://github.com/yourname/ComfyUI_dct.git
   

2. Install the Python dependency:

   pip install scipy
   

3. Restart ComfyUI. The three nodes appear under the DCT category in the node search menu.

Requirements

Package	Version
scipy	≥ 1.4.0
Pillow	any recent
torch	(provided by ComfyUI)
numpy	(provided by ComfyUI)

---

Technical Details

• Transform: 2-D DCT Type-II, separable, applied independently per colour channel.
• Normalisation: norm='ortho' — the forward and inverse transforms are exact inverses of each other.
• Mask resolution: The paint canvas uses a 512 × 512 internal raster. At execution time the mask is bilinearly resampled to match the actual spectrum dimensions.
• Data flow: DCT_SPECTRUM is a standard torch.Tensor [B, H, W, C] carrying real-valued DCT coefficients; it is not clamped to [0, 1].