🚀 Next-gen semantic compression

Preserve Identity
Compress Everything

HELIX stores only what defines uniqueness. AI regenerates the rest. Up to 2-5x semantic compression with identity preservation.

Try Demo
2-5x
Compression
0.85+
Identity Match
3
Modalities

What is HELIX?

A multimodal semantic compression and regeneration system. A new file format (.hlx) that stores identity-critical information and structural constraints rather than raw data.

Not an AI Model

HELIX is a representation system, compiler, and constraint engine. It controls AI, doesn't replace it.

The .hlx Format

A compiled specification for regeneration, not a media container. Stores instructions, not content.

AI as Executor

External AI models serve as deterministic regenerators under strict constraint supervision.

Entropy Separation

Compression by discarding predictable entropy, not by squeezing bits harder.

❌ Traditional Compression

  • • Stores all data with pattern-based reduction
  • • Decoder reconstructs mechanically
  • • Limited by signal redundancy
  • • Asymptotic compression limits

✓ HELIX Compression

  • • Stores only identity + structure
  • • Decoder regenerates intelligently
  • • Limited by semantic predictability
  • • Leverages model priors, not just patterns

How HELIX Works

Two pipelines transform media into compact blueprints and back.

Encoding

Media → .hlx
Normalize

Establish canonical representation

01
Detect Saliency

Find identity-bearing regions

02
Extract Anchors

Isolate irreducible features

03
Build Mesh

Encode geometric relationships

04
Set Constraints

Define absolute invariants

05
Assemble .hlx

Serialize encrypted blueprint

06

Materialization

.hlx → Media
Validate

Verify integrity and hashes

01
Bind Executor

Prepare Gemini with constraints

02
Regenerate

AI generates within bounds

03
Verify

Ensure identity preservation

04

Key insight: Anchors are stitched exactly — AI only regenerates the predictable background.

System Modules

Each module plays a critical role in the HELIX pipeline. Click to explore.

HLX Codec

hlx_codec.py

AES-256-GCM encryption + HMAC-SHA256 signing. Binary format only HELIX can read.

Tamper detectionMagic header verificationVersion control

The codec ensures that .hlx files are completely proprietary. Any attempt to modify the file triggers a HLXTamperingError on load.

Anchor Extractor

gemini_anchors.py

Gemini-powered identity extraction with 0.85 confidence threshold.

Face componentsText OCRObject detectionSHA-256 hashing

Uses multi-model cascade (Gemini 2.0 → 1.5 Flash) for robustness. Extracts hierarchical anchors.

Mesh Builder

mesh_builder.py

Constructs structural relationships: distances, angles, ratios.

Normalized coordinatesFacial proportionsAdjacency graphs

Encodes geometric constraints without storing content. Prevents AI from warping proportions.

Blueprint Validator

blueprint_validator.py

Enforces schema integrity before materialization.

Version checkHash verificationConflict detection

Checks schema version compatibility, verifies anchor hashes, detects freedom field conflicts.

Materializer

materializer.py

Multi-layer regeneration: stitching, enhancement layers, optional AI refinement.

Anchor stitchingCLAHE contrastEdge sharpeningDetail injection

First stitches anchor crops onto canvas, then applies deterministic enhancement layers (sharpening, contrast, denoising). Optional Gemini AI refinement when API available.

Verification Loop

verification.py

Post-generation validation ensuring identity preservation.

Drift <2%Hash matchingSSIM scoring

Re-detects anchors in output image, compares positions to originals. Allows <2% position drift.

How HELIX Works

Semantic compression that preserves meaning, not pixels

2.5 MB
250 KB
.hlx
AI Restored
Original
Compress
Store
Materialize
10x
Average Compression Ratio

Try HELIX

Upload a test image to see compression in action.

Drop your image here

or click to browse

Images

Login required to process files