Checkerboard Artifact
Definition
A grid-like pattern visible in the Fourier power spectrum of GAN outputs, caused by transposed convolution or bilinear-upsampling operations used to increase image resolution in the generator. The pattern is often invisible to the eye but detectable by spectral analysis.
- Cause
- Transposed convolution or bilinear upsampling in GAN generator
- Domain
- Fourier power spectrum, not spatial image
- Visibility
- Usually invisible to the naked eye
- Detection method
- Spectral analysis of frequency-domain periodicity
Common questions
Why is spectral analysis needed instead of just viewing the image?+
The grid pattern arises from a periodic upsampling operation and appears as regularly spaced peaks in the frequency domain, but at spatial resolution it blends into normal image texture, so it only becomes visible after a Fourier transform.
Does the checkerboard artifact appear in diffusion-generated images too?+
It is characteristic of GAN architectures using transposed convolution or naive upsampling. Diffusion models use different sampling operations and generally produce a different, weaker or absent version of this specific periodic signature, which is one way examiners distinguish generator families.
Related terms
- Autoencoder
- A neural network that compresses an input into a compact latent representation (encoder) then reconstructs it (decoder). Face-swap pipelines train autoencoders with...
- Blending Mask
- In face-swap pipelines, a pixel-level mask that defines the face region to be composited onto the target frame. Imperfect masks leave boundary...
- Diffusion Model
- A generative neural network architecture (Ho et al., 2020; Stable Diffusion, Rombach et al., 2022) that learns to reverse a noise-addition process...
- Generative Adversarial Network (GAN)
- A machine learning architecture consisting of two networks trained simultaneously: a generator that produces synthetic images and a discriminator that classifies images...
- Latent Space
- The compressed, lower-dimensional representation of data learned by a neural network's internal layers. Generative models sample from or navigate this space to...