Kelin9BT vs ErnieIT vs ZIT (FFT Analysis of Artifacts)

A detailed technical comparison of three leading AI image generation models has gone viral, providing a quantitative look at visual quality through artifact analysis. Reddit user ZerOne82 conducted a controlled test pitting Stability AI's Flux 2 Klein 9B Turbo against Baidu's Ernie Image Turbo and another model called Z-Image Turbo. Using the prompt "extreme close-up of a woman with long brunette and blonde hair covering half her face. she is holding a cardboard sign with text 'artifacts'," the test employed Fast Fourier Transform (FFT) analysis—a mathematical technique for detecting repeating patterns and imperfections—to objectively measure output cleanliness.

The results were striking. The FFT graphs revealed that Klein 9B Turbo produced the cleanest spectral output with minimal artifacts, while Ernie Image Turbo showed the "dirtiest" output with clear diagonal artifacts visible in the frequency domain. According to the analysis, these artifacts in Ernie's output were particularly noticeable in realistic renders, especially in hair details, and proved persistent—"no amount of tweaking with different samplers and steps could remove them." The test used consistent parameters (848x1264 resolution, Euler-A sampler, beta scheduler) at both 4 and 8 steps, highlighting fundamental differences in how these models handle image synthesis at a mathematical level.

This analysis moves beyond subjective visual comparison to provide developers and researchers with an objective, data-driven method for evaluating image generation quality. For professionals working with generative AI, understanding these artifact patterns is crucial for selecting models for production use, especially in applications requiring high-fidelity visual outputs where such imperfections could undermine quality.