PUCE Image
Neural image codec with integrated post-quantum encryption. Next-generation compression for critical visual data.
State-of-the-art neural compression
Four pillars that define the most advanced image codec on the market.
CVPR 2025 Neural Compression
Neural model with 45.6M parameters. Internally measured BD-Rate (PSNR, Kodak): −72% vs JPEG, −53% vs WebP, −42% vs AVIF, −30% vs VVC intra.
Post-quantum encryption
AES-256-GCM with ML-KEM-768 key exchange integrated in the bitstream. Each tile can have an independent key.
Tiled processing
Large images split into tiles for parallel GPU processing. Enables arbitrary resolution without memory limits.
Distilled model: PUCEImageFast
Compact version with only 2M parameters (96% reduction). Ideal for edge devices and real-time inference.
BD-Rate Benchmarks
BD-Rate savings of PUCE Image vs each codec. Negative values = fewer bits for the same quality.
| Reference codec | BD-Rate |
|---|---|
| JPEG | -72.3% |
| WebP | -53.0% |
| AVIF | -41.6% |
| VVC Intra | -30.0% |
Pipeline architecture
Six stages — from content analysis to post-quantum encryption.
Content analysis
Automatic image type classification (photo, graphic, text, medical) for adaptive routing
Neural encoder
Proprietary neural encoder — 8.5M parameters, learned non-linear transform with multi-scale attention
Spatial prior
13.7M parameters — autoregressive model with spatial context for efficient entropy coding
Hyperprior
1.5M parameters — captures global latent structure for distribution estimation
Neural decoder
Proprietary neural decoder — 14.9M parameters, reconstruction with attention and skip connections
PQ encryption
AES-256-GCM with ML-KEM-768 key exchange — native post-quantum protection in bitstream
Format specification
| Extension | .puce-img |
| Magic bytes | 0x50 0x55 0x43 0x45 ("PUCE") |
| Encryption | AES-256-GCM + ML-KEM-768 |
| Entropy coding | Arithmetic (learned priors) |
| Bit depth | 8/10/12-bit |
| Color space | RGB / YCbCr 4:4:4 / 4:2:0 |
| Tiling | Adaptive (256×256 to 1024×1024) |
Use cases
Secure medical image storage (DICOM, digital pathology)
Professional photography with neural compression superior to HEIF/AVIF
Satellite and geospatial imagery with military-grade encryption
E-commerce — catalogs with 60%+ reduction vs JPEG with no visible loss
Long-term digital archive with quantum protection
Classified government applications (CNSA 2.0 compliant)