Seamless Texture Pre-Processor

The most common reason a "seamless" texture still shows a seam has nothing to do with detail - it's brightness. A photo or render lit from one side has a slow gradient across it, and when you wrap it into a tile, the bright edge butts against the dark edge. No amount of edge blending fixes that; the two sides simply don't match in tone.

Seamless Texture Pre-Processor is the pre-flight step that tries to stop you from failing before you start. It does two things: crops junk off the edges, and flattens large-scale lighting gradients so the tileable result doesn't have a visible brightness band at the seam.

How it works

Two stages, both pure torch, both deterministic:

1. Pre-crop. Cuts pre_crop_left/top/bottom/right pixels off each edge. Handy for removing a vignette, a sky gradient, stray UI text, or content you don't want wrapping around the tile.
2. Lighting correction. This is the clever bit. It computes luminance, blurs it with a periodic (wrap-around) Gaussian - so the blur itself tiles - then scales each pixel so the blurred luminance is pulled toward the image's mean. The scale is clamped to 0.25–4.0 and blended in by the intensity. Net effect: broad lighting changes flatten, while fine detail keeps its local contrast.

The blur radius is normalized to image size - sigma ≈ (min(w,h)/256) × radius - so the default behaves consistently whether you feed it 512px or 2048px. And note the default lighting_correction_intensity is 65, not 0: this node is actively correcting out of the box.

The inputs that matter

• lighting_correction_intensity (0–100, default 65) - how hard to flatten. 0 disables it entirely if you only want the crop.
• lighting_correction_radius (1–20, default 5) - the scale of gradient it targets. Small radius corrects fine variations (and starts washing out local contrast); big radius only touches broad gradients. You'll rarely need to move it.
• pre_crop_* - leave at 0 unless you actually have border junk.

The output is a single IMAGE (image), and it's designed to feed straight into Seamless Texture Radial Mask or Half Shift. The README's recommended chain, in fact, is: generate a texture (it suggests Flux Klein 4b or Z-Image Turbo with a circular VAE decode from spinagon's ComfyUI-seamless-tiling), push it through this pre-processor, then through the radial mask. The pre-processor exists to catch exactly the uneven lighting those generation paths leave behind.

Installation

Identical to the rest of the pack - no models, no extra dependencies beyond torch:

cd ComfyUI/custom_nodes
git clone https://github.com/SparknightLLC/ComfyUI-MakeSeamlessTexture

or search "ComfyUI-MakeSeamlessTexture" in ComfyUI Manager. Restart after installing.

Common issues

• Correction too weak or too strong. Push intensity to 100 for a stubborn gradient, or drop to 0 if it's flattening the image into a gray mush. There's no middle setting that fits every source.
• Radius wrong for your image. If you see the result look "veiny" or locally flat, the radius is too small - the correction is tracking local features instead of the gradient. Crank it up.
• Cropping different-sized inputs. If you batch images that end up different sizes after the crop, the stacked output can fail. Feed it one image (or same-sized ones).