Skip to content
LeChiffreVol2Public

About

Inference for Generative Foundation Models for Earth Observation. Python, CLI and local API.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

AUREON — Generative Foundation Models for Earth Observation

Generative Foundation Models for Earth Observation.
Specialized inference. Python, CLI, and a local API.

CPU and static checks Release v0.1.0 alpha Python 3.12 License Apache-2.0

Website · Documentation · Quickstart · Models · Download · Contribute


Public alpha v0.1.0. Install one Engine for Python, CLI and local HTTP. Start with TerraMind generation and your own Sentinel-2 imagery. The current GitHub download is Best e09e710c; PyPI and hosted inference are not open.

AUREON is an open-source inference engine built for Generative Foundation Models for Earth Observation, with support for the broader Earth foundation model family. Start with TerraMind: turn a prepared Sentinel-2 L2A crop into synthetic S1GRD VV/VH imagery, delivered as a georeferenced Cloud Optimized GeoTIFF (COG) with provenance. KAO-TerraMind adds dense-prediction workflows through the same Engine.

Use Python, CLI or an authenticated local API. The published v0.1.0 library runs on your own machine; the landing page describes the wider platform direction.

Built on TerraTorch and PyTorch, AUREON brings input validation, resident inference and geospatial result delivery into one workflow.

The package includes bounded stage execution and remote replica contracts: typed job handles, resource credits, tenant fairness, durable leases and verified artifact delivery. Native hardware capability and numerical qualification are separate. These interfaces do not claim measured multi-GPU/multi-host scaling.

Whole scenes and repeated jobs

The integrated workflow connects the unchanged numerical core to whole-scene delivery, resident sessions and bounded repeated jobs. 224x224 is a native tile size, not a scene-size limit. Budgets, masks, coordinate seeds, cancellation, committed COG/provenance and recovery are explicit. Historical aliases remain compatible; no second package is needed.

Acceptance receipts bind exact wheel bytes and tested routes. Historical performance figures retain their original experiment scope. Exact numerical parity does not establish seamless physical quality, long-load stability or physical multi-GPU speed.

Read how we evaluate quality and efficiency for measurement boundaries, numerical quality, usability and remaining qualification. Historical competitor percentages stay attached to the original measured Best 0fa5ef2b.

Why AUREON?

Built for What you get
Geospatial inputs and outputs Explicit band order, radiometry, grids and masks; georeferenced results you can inspect in GIS tools.
Repeated model execution Resident sessions, exact eligible encoder reuse and guarded CUDA Graph execution for supported profiles.
Traceable results Model and preprocessing identities, FP32 reference checks, provenance and verified artifact delivery.
A workflow you can own Python, CLI and authenticated local HTTP, with separate locked dependency profiles and explicit model execution.

Supported workflows

The package includes prepared-scene generation, DataLoader integration and read-only runtime diagnostics. Consult that guide for limits and artifact qualification; the current Best download includes these interfaces; original alpha receipts retain their earlier artifact scope.

Model Task Input → output Start here
TerraMind 1.0 base Cross-modal generation All-valid 224 × 224 S2L2A → synthetic S1GRD VV/VH First generation
KAO-TerraMind B Segmentation Twelve-band Sentinel-2 L2A → class-map and validity-mask COGs First scene

KAO-TerraMind A/C also have implemented full-reference routes with earlier-artifact evidence. See the support matrix for exact profiles, hardware and qualification. Generated S1GRD is synthetic imagery, not an observed SAR measurement.

Install

Linux x86_64 or WSL2 · Python 3.12 · NVIDIA GPU for model execution

Install the current Best wheel using the single distribution guide, then follow input/asset preparation and generation. The model/tokenizer and user imagery remain explicit assets and are not bundled. Download Best v0.1.0. The wheel SHA-256 is e09e710c9e9ace17b7bf2214d8c9eff2c4ed605a7d45e8c4e649b672f6f88a64. Original alpha assets remain under historical names on the same release; verify hashes rather than version alone.

python -m pip install "https://github.com/LeChiffreVol2/aureon/releases/download/v0.1.0/aureon_inference-0.1.0-py3-none-any.whl#sha256=e09e710c9e9ace17b7bf2214d8c9eff2c4ed605a7d45e8c4e649b672f6f88a64"

For an isolated, reproducible setup from this checkout, use:

python3.12 tools/install.py --execute --prefix "$HOME/aureon-env"

The installer verifies the unchanged release wheel, installs hash-locked dependencies, checks installed payloads and runs the CPU preparation walkthrough. Use --plan to preview. First-use instructions and feedback cover offline installation, input preparation and recovery.

Use a fresh Linux/WSL Python 3.12 environment. For the tested pinned stack, use the matching source archive locks and release instructions. An existing installation also named 0.1.0 may still contain the original wheel.

After installation, inspect your environment without loading a model:

aureon-engine models
aureon-engine assets --model terramind-v1-base
aureon-engine check --environment generation

A first look at the API

After creating your TerraMind config, run the same first-generation workflow from Python. Choose either this example or the CLI job in the guide; both execute a model job.

from pathlib import Path
from aureon.api import GenerationRequest, run_model

work = Path.home() / "aureon-work"
request = GenerationRequest(
    config=str(Path.home() / "aureon-config/generation.json"),
    output=str(work / "generation-python-first"),  # Must not exist yet.
    job_id="first",
    attempt_id="one",
    reuse="auto",
    denoiser="graph",
)
result = run_model(request, execute_model=True)

print(result["job"]["cog"]["file"]["path"])
print(result["job"]["provenance"]["path"])
print(result["cleanup"]["passed"])

execute_model=True explicitly starts inference. For a CPU-only installation walkthrough without imagery or weights, use the optional preparation demo. It checks the input pipeline with artificial pixels; it does not generate a model prediction.

Go further: Generation and resident sessions · CLI quickstart · Local HTTP API

How it fits together

flowchart LR
    A[Your Sentinel-2 imagery] --> B[Prepare bands, grid and masks]
    B --> C[AUREON Engine]
    D[Pinned model assets] --> C
    C --> E[Synthetic S1GRD or class-map COGs]
    C --> F[Masks and provenance]
Loading

Choose Python, CLI or local HTTP for the same execution layer. The architecture guide explains the interfaces and result lifecycle.

Release status

The current local v0.1.0 distribution is the integrated best-core Engine described in the installation guide. Its delivery receipt identifies one wheel/source pair, installed checks and the exact scope of model acceptance. There is no current alternative feature build.

The original published alpha remains unchanged. The original performance evidence also remains bound to its measured artifact, configuration and workload. Neither historical release files nor historical measurements are overwritten. No hosted API, production SLA or physical multi-GPU qualification is announced.

Contributing

Bug reports, clearer examples, documentation fixes and focused Engine improvements are welcome. Start with the contributor guide for the repository map and CPU checks, or open a bug report or feature request. For installation problems, see troubleshooting.

License and acknowledgments

AUREON Foundation by Norvique is a project of Norvique Company. AUREON-owned contributions are licensed under Apache-2.0. We acknowledge the work of KAO-TerraMind, TerraMind, TerraTorch and PyTorch. See NOTICE, third-party attribution and licensing scope for details. Model weights and datasets retain their respective terms.

Contact: norviccompanylimited@gmail.com. Report vulnerabilities privately through SECURITY.md.

About

Inference for Generative Foundation Models for Earth Observation. Python, CLI and local API.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages