I compared the reconstruction quality of the latest VAE models (Focusing on small faces). Here are the results!

A developer's viral benchmark test has identified a new leader in image reconstruction quality for AI models, specifically for the challenging task of rendering small faces. The test, conducted for a face-editing project, pitted the latest VAE (Variational Autoencoder) decoders—critical components that convert AI latents into final images—against each other. The Flux2 Klein VAE, developed by the Black Forest Labs team behind the Flux models, emerged as the definitive winner. It demonstrated superior handling of micro-details and textures in small facial features, a task where other models faltered. The test included Zimage (the VAE for Flux1), the QwenImage VAE from Alibaba's Qwen team, and standard Stable Diffusion and SDXL VAEs for baseline comparison.

The results highlight a significant technical achievement. The Flux2 Klein VAE's performance suggests a "massive amount of effort" was invested in its training, according to the tester. In contrast, the QwenImage VAE showed noticeable artifacts and struggles with small-face reconstruction. The Zimage VAE held its own, indicating solid performance from the previous generation. This comparison is vital because the VAE's quality directly impacts the final output of text-to-image and image-editing workflows; a poor decoder can ruin the work of a powerful diffusion model.

For professionals in AI image generation, editing, and post-production, this benchmark provides actionable data. Choosing the Flux2 Klein VAE for pipelines requiring high-fidelity human faces, especially in complex scenes with multiple small subjects, could lead to significantly better outputs without changing the core diffusion model. It underscores the importance of component-level optimization in the AI image stack and signals where competing teams like Stability AI (SDXL) and Alibaba (Qwen) need to focus improvements.