CAESAR software architecture
The architecture figure replaces the earlier workflow diagram. It shows model registration and installation, compression, local reconstruction, residual correction, serialized metadata, and the corresponding decompression path.

Download the vector PDF ·
Download the editable SVG ·
Download the Python generator
Compression reads left to right, and decompression reads right to left. Enclosing
panels group each phase. Matching component colors show related operations:
blue for neural models, orange for correction, Gator Green (#22884C) for rANS,
and white for data and metadata. Blue (#1B4FA8) and orange (#E8500A) match
the README. Solid arrows show data and metadata; dashed arrows show installed
model dependencies. Both paths reuse the matching installation’s packages and
probability tables.
The version-1 CAESAR serializer is implemented: serialize/deserialize preserve
all result fields in a buffer, while save/load provide single-file storage.
Registered model weights and entropy tables remain installed separately; trained
NGLR correction weights, when used, travel with the result. See
the public API.
Regenerate the figure
Install Matplotlib in your Python environment, then run from the repository root:
python3 docs/generate_architecture.py
The generator writes PDF, SVG, and an 18.5 × 10.5 inch PNG at 300 dpi
(5550 × 3150 pixels). All labels are bold; font_scale controls their sizes.
It checks box text for horizontal overflow and title/detail overlap.
The figure summarizes component relationships rather than every execution branch. Constant/filtered blocks can bypass neural coding. Input conversion currently selects variable 0 for 5D tensors. The entropy-decoding block includes the hyperprior operations needed to recover latent symbols. The CLI’s legacy three-file layout is distinct from the version-1 buffer and single-file API; external ADIOS2 migration is tracked separately.