APEX: Assumption-free Projection-based Embedding eXamination Metric for Image Quality Assessment

Researchers introduce APEX, a novel image quality assessment metric that addresses fundamental limitations in existing evaluation methods like FID by using Sliced Wasserstein Distance and modern foundation models (CLIP, DINOv2) as embedding-agnostic feature extractors. The framework eliminates parametric assumptions while maintaining scalability to high-dimensional spaces, demonstrating superior robustness and stability across datasets.

APEX represents a meaningful advancement in how researchers evaluate generative model outputs, a critical capability as AI-generated images increasingly approach indistinguishable quality from authentic content. Traditional metrics like FID rely on fixed feature representations from outdated neural networks, creating a 'closed-vocabulary bottleneck' that prevents accurate assessment of novel visual qualities. This limitation becomes problematic when evaluating cutting-edge generative models that produce content beyond what older feature extractors were trained to recognize.

The research addresses this by abandoning rigid parametric assumptions—mathematical formulations that impose constraints on how similarity is measured. Instead, APEX leverages Sliced Wasserstein Distance, a mathematically grounded approach from optimal transport theory that doesn't require predefined assumptions about data distribution. By integrating modern foundation models like CLIP and DINOv2, the framework gains access to richer, more generalizable feature representations trained on diverse, large-scale datasets.

The implications extend across the generative AI industry. Better evaluation metrics enable faster iteration cycles for model developers, more reliable comparisons between competing approaches, and increased confidence in deployment decisions for applications requiring certified output quality. The demonstrated stability across out-of-domain datasets suggests APEX could become a standard reference metric, similar to how FID currently dominates industry benchmarking despite its known limitations.

Developers building image generation systems, from text-to-image models to video synthesis pipelines, would benefit from more trustworthy evaluation tools. The research validates that modern foundation models can replace outdated feature extractors while theoretical grounding prevents overfitting to specific assumptions. Future work may extend this framework to video assessment or multimodal evaluation scenarios.

• →APEX eliminates parametric assumptions from image quality assessment using Sliced Wasserstein Distance for mathematically grounded similarity measurement.
• →Integration of CLIP and DINOv2 foundation models overcomes the closed-vocabulary bottleneck limiting older metrics like FID.
• →Framework demonstrates superior robustness to visual degradations and maintains stability when evaluating out-of-domain datasets.
• →Research addresses a critical bottleneck in generative AI development by enabling more accurate evaluation of model outputs.
• →Potential to become industry standard for image quality assessment, replacing FID-based benchmarking workflows.