Technology · Visual Structure & Inspection Layer
ByteSight
Deterministic visual-structure engine with an operational segmentation and teaching workbench.
Mission
Vision that can show its work.
Deterministic visual-structure engine with an operational segmentation and teaching workbench, editable regions, structural labeling, saved-work integrity, and canonical partition export. Fine facial segmentation and production workflow refinement remain.
Where a typical classifier returns a confidence score with no explanation, ByteSight represents canonical regions, boundaries, and closure state explicitly, with a desktop workbench that lets a person inspect and correct why a region was identified the way it was. This is implemented and internally tested today — see Current Capabilities below for what's proven and Validation and Evidence for exact scope and known limitations, including a documented gap on close-up portrait photos.
Current Capabilities
What exists today.
Deterministic image partitioning into discrete regions connected by an explicit adjacency and parent/child graph, with hash-verified reproducibility across runs and sessions — including one real non-determinism bug that was found and fixed during testing
Region-boundary closure validated for exact, byte-identical reconstruction; construction is refused rather than silently degraded if this check fails
A reproducible, bijective canonical partition export format (structured manifest plus raster) for general, non-facial image partitions, exercised by the desktop workbench
Ten general object/material roles (e.g. skin, hair, cloth, background) classified per region and used at render time
Saved settings and presets verified to round-trip to identical values on reload, tested field-by-field
An interactive native desktop workbench for layer editing, pixel-level region correction (in-app painting or imported mask), and an annotation review lifecycle (draft/reviewed/accepted/rejected/superseded), with atomic, audit-trailed saves
Manual authoring and exhaustive validation of six facial-region categories (eyes, eyebrows, nose, mouth, hairline, facial hair) via in-app painting or imported masks — always human-marked, never automatically detected
In Development
Active work.
- Relationally positioned pixel evidence (describing a pixel by its exact structural relationship — adjacent, contained, directional offset — to other known points in the same image); implemented and passing tests as of the current build, but not yet in the stable, committed proof corpus
- Wiring the 13-category facial-annotation taxonomy into the renderer — it currently exists for manual authoring and validation only, and is not yet consulted when rendering output
- An output-composition layer for Cordel-style profile photos (crop, framing, sizing) — not yet built
- Automatic-mode testing against real uploaded photos — current automatic-path testing uses synthetic images; a real close-up portrait test recorded a known failure (zero facial features detected), the same photo shape a dating-profile picture would use
End-Game Functionality Goals
The target architecture — not the current system.
Everything below describes what ByteSight is designed to become. It is not a claim that this exists today.
- Full Cordel Connect photo-transformation readiness, including tested performance on real close-up portrait photos
- The facial-region taxonomy wired into the renderer rather than annotation-only
- Relationally positioned pixel evidence promoted from active development into the stable proof corpus
- A production adapter feeding Cordel Connect's cartoonized-profile feature and Deep Kore ingest
Validation and Evidence
Proof posture.
- Internal validation, 2026-07-31 (commit 654de66, local C++/CMake build): deterministic partitioning, boundary closure, canonical export, saved-work round-trip, and manual facial-region authoring are covered by the automated test suite and dedicated proof artifacts, including one documented case of a real non-determinism bug found and fixed.
- No biometric face-matching, identity-recognition, or machine-learning training exists anywhere in this system — confirmed by an explicit in-code architectural disclaimer and by direct inspection; facial regions are always human-marked, never automatically detected.
- A real test on a close-up portrait photo recorded a known failure (zero facial features detected) — this is a documented current limitation, not a resolved capability.
- Not independently verified by a party outside ByteLite LLC.
- Not a claim of complete Cordel Connect photo-transformation readiness — the output-composition layer for profile-photo delivery is not yet built.
What this is not.
Integration
How it connects.
Receives
- Raw images and visual media
Produces
- Region/boundary structural output and canonical partition exports
- Provenance-tagged, audit-trailed correction records
Consumed by
- Deep Kore (planned ingest)
- Cordel Connect (contract defined for a photo/cartoonizer adapter; not yet integrated end-to-end)
Availability and Commercial Access
Related