FRAPPE: Full Input, Residual Output Autoencoding with Projection Pursuit Encoder

Researchers introduced FRAPPE, an autoencoding framework for learned image compression that predicts residuals from full inputs using a projection pursuit encoder.

Applied to FRAPPE-Image, it enables variable-rate RGB coding with zero-overhead rate adaptation and parallel encoding.

The system delivers faster CPU-based compression and real-time 1080p encoding on CPU.

At very low bitrates, it achieved better perceptual quality than AVIF.

The work makes learned codecs more practical for CPU-only and low-power devices.