New VHS verifier reduces AI image generation costs by skipping pixel decoding and re-encoding steps

Researchers propose Verifier on Hidden States (VHS), which operates directly on Diffusion Transformer (DiT) intermediate representations instead of requiring full pixel decoding

VHS eliminates redundant operations by analyzing generator features in latent space, avoiding costly re-encoding into visual embedding space

The approach matches or exceeds performance of Multimodal Large Language Model (MLLM) verifiers while significantly reducing per-candidate verification computational cost

Addresses inference-time scaling bottleneck where traditional verifiers decode candidates to pixel space, creating inefficient workflows in autoencoder-based diffusion pipelines