ComfyUI-DeGrid Introduction
ComfyUI-DeGrid adds one ComfyUI node, VAE DeGrid (Nyquist Notch). It removes a faint, repeating 2-pixel grid that can appear when certain VAEs decode an image. The issue affects Qwen Image and Qwen Image 2.1 outputs, and can also appear with the Wan 2.1 VAE and workflows built on these VAEs, including Krea2 and Anima.
You may not notice the grid at normal viewing size. It is often easiest to see in smooth, dark areas, where a regular pattern stands out against an otherwise even surface. Sharpening and upscaling can make it much more obvious. DeGrid is intended to remove the pattern before those later steps amplify it.
The node measures each image and, by default, adjusts its correction automatically. If it finds no grid, it leaves the image untouched.
How ComfyUI-DeGrid Works
Think of the unwanted grid as a very faint pattern stamped onto every other row and column of an image. Real image details—such as hair, fabric, and brush-like texture—vary from place to place. The VAE artifact is different: its 2-pixel pattern stays aligned across the frame.
DeGrid uses that regularity in two ways:
- It checks whether the grid is present. The node compares repeating pixel positions across the image. This helps it distinguish a consistent lattice from ordinary fine detail.
- It removes the pattern while limiting changes to real detail. Its filter targets 2-pixel checkerboard, vertical-stripe, and horizontal-stripe patterns. A correction limit helps protect stronger edges and textures.
In auto mode, the node estimates an appropriate correction limit separately for each image. This is useful when you switch between images, VAEs, or workflows without wanting to tune a setting each time.
Placement is important: connect VAE Decode → VAE DeGrid → sharpening, resizing, restoration, or upscaling. If you resize or upscale first, the original lattice may be altered or amplified into new-looking detail that DeGrid can no longer identify and remove reliably.
The approach is based on a GLSL notch-filter technique shared on r/StableDiffusion. The author implemented this node in PyTorch and added correction limits, automatic calibration, and grid detection.
ComfyUI-DeGrid Features
Automatic filtering and a readable status
For most images, leave mode set to auto. After processing, the node displays a status line such as:
grid 2.10/255 (checker) — removed (limit 0.019 auto) · edges protected: 1.2%
Here is how to read it:
grid 2.10/255is the measured strength of the repeating pattern. The value uses the familiar 0–255 image scale; a small number can still become noticeable after sharpening.checkernames the strongest detected pattern. You may instead seeV-stripeorH-stripe.limit 0.019 autoshows the correction cap chosen for this image.edges protected: 1.2%shows the portion of pixels where the correction reached that cap, helping preserve stronger image detail. A few percent is normal.
When the measurement is below the node’s detection threshold, the default behavior is to pass the image through unchanged. The status may then look like:
grid 0.16/255 — none detected, passed through untouched · edges protected: 0.0%
Controls
| Setting | Default | What it means for you |
|---|---|---|
enabled | On | Turn it off to pass the image through untouched. This is useful for a quick before-and-after comparison. |
mode | auto | Automatically measures each image and selects a correction limit. Choose manual only if you can see that automatic filtering leaves a grid behind or softens fine detail. |
limit | 0.02 | Used only in manual mode. It caps how much an individual pixel can be corrected, on a 0–1 scale. A lower value preserves more tiny texture but may leave some grid; a higher value removes more grid but may soften pores or fabric weave. |
skip_when_clean | On | Leaves an image completely untouched when no grid is detected. Turn it off only if you deliberately want to run the filter despite that result. |
grid_gain | 10 | Brightens the removed_grid preview so you can inspect what the node removed. It does not change the cleaned image. |
grid_view | 4x zoom | Controls the preview framing: full frame, 4x zoom, or 8x zoom. Zoom modes show a magnified center crop of the removed pattern; they do not crop the cleaned image. |
The removed_grid preview
This preview shows the correction rather than your finished artwork. At thumbnail size, a 2-pixel lattice may look like gray noise because the preview is too small to display every pixel separately. To inspect it, use 4x zoom or 8x zoom, or open the preview at 100% zoom or higher.
- A fine, regular pattern: The node is removing the kind of lattice it targets.
- Almost flat gray: There may be little grid to remove; check the status line.
- Clearly recognizable faces, edges, or fabric detail: The correction may be too strong. Try
manualmode with a lowerlimit.
A faint outline of the subject can be normal, because the artifact may be stronger over detailed areas. Distinct, recognizable detail in the removed preview is a better reason to reduce the limit.
ComfyUI-DeGrid Models
DeGrid does not provide a collection of models to select from. It is a single image-processing node designed to address an artifact associated with certain VAEs.
The documented affected VAEs and related workflows include Qwen Image, Qwen Image 2.1, Wan 2.1, Krea2, and Anima. Qwen Image 2.1 uses a different VAE architecture from the earlier Qwen Image / Wan 2.1 VAE, but its observed artifact is still a 2-pixel lattice. You do not need to choose a different DeGrid model for it; the node measures the decoded image.
What’s New with ComfyUI-DeGrid
2026-09-20
- More accurate grid detection: The node now measures the repeating lattice itself rather than treating an image’s general fine detail as evidence of a grid. This makes its status more meaningful, particularly for highly detailed artwork.
- Clean-image pass-through: The new
skip_when_cleansetting is on by default. When the node detects no grid, it leaves the image unchanged instead of applying an unnecessary filter. - Clearer status messages: The status names the dominant checkerboard or stripe orientation and no longer says a grid was removed when none was detected.
- Qwen Image 2.1 confirmed: Native decodes were found to contain the same type of 2-pixel artifact, so the existing filter applies to them as well.
The grid-removal filter, its correction cap, and automatic limit calculation did not change in this update. Images that genuinely contain the grid receive the same filtering as before.
2026-07-04
- Initial release of the DeGrid node.
Troubleshooting ComfyUI-DeGrid
The grid is still visible after filtering
- Confirm the node is connected directly after
VAE Decode, before resizing, sharpening, or upscaling. - Read the status line to see whether a grid was detected.
- If a grid was detected but remains visible, switch
modetomanual. - Raise
limitgradually toward0.03–0.04, comparing the cleaned image andremoved_gridpreview as you go. Stop if real texture begins to soften.
Fine texture looks softer
- Switch
modetomanual. - Lower
limitgradually toward0.01. - Compare details such as pores, hair, and fabric with
enabledon and off. A lower limit protects more texture, though it may leave some grid.
The status says “none detected,” but I can see a pattern
- Check whether the pattern is already present immediately after
VAE Decode. - If you placed a resize between decoding and DeGrid, move DeGrid before the resize; resizing can disrupt the lattice the node looks for.
- If the pattern appears only after a sharpener or upscaler, investigate that later step. DeGrid targets the VAE’s original 2-pixel artifact, not every grid-like pattern a workflow might produce.
The removed_grid preview looks gray
- Check the status line:
none detectedmay simply mean there was little or nothing to remove. - Set
grid_viewto4x zoomor8x zoom, or open the preview at 100% zoom or higher. - Increase
grid_gainif you need a brighter preview. This will not alter your finished image.
Frequently asked questions
Do I need to adjust the settings for every image?
Usually not. auto mode measures each image and sets the correction limit for you.
Will DeGrid change images that do not contain the VAE grid?
With skip_when_clean on, images below the grid-detection threshold pass through untouched. If you turn that setting off, the filter runs even when the status says none detected.
Why should I run it before an upscaler or restorer?
Those tools can treat the regular lattice as image detail and strengthen or reconstruct it. Removing the artifact immediately after decoding gives later steps a cleaner image to work with.
Can I use ComfyUI-DeGrid alongside ComfyUI-SaveSimple?
No. ComfyUI-SaveSimple includes the same node under the same identifier, VAEDeGrid, so using both creates a node-ID collision. Choose DeGrid if you only want this grid fix, or SaveSimple if you use that wider suite.
Learn More about ComfyUI-DeGrid
- ComfyUI-DeGrid project repository — the source for the node and its project documentation.
- Discussion of the 2-pixel VAE grid and notch-filter approach — background on the artifact and the filtering idea that inspired the node.
ComfyUI-DeGrid is licensed under Apache-2.0.
