GAR Introduces Generation-Aware Reconstruction: Aligning Decoder with Latent Distribution Yields 0.99 Spearman Correlation

Addressing the disconnect between reconstruction metrics and generation quality in image models, researchers propose the Generation-Aware Reconstruction (GAR) framework. By injecting noise into encoded latents and simulating denoising, GAR creates a continuous evaluation path from reconstruction to generation. Experiments show that the resulting GAR-FID metric correlates strongly with final generation quality, achieving a Spearman correlation of 0.99 in specific settings, effectively resolving the weak or negative correlation seen with traditional rFID.