v0.3.0 - Moon and Twilight LDM Models

What's New in v0.3.0

This release adds two new pre-trained Latent Diffusion Models

New Models

• ldm_moon: Moon-contaminated sky spectra

  • 6 conditioning features: OBSALT, TRANSPARENCY_GFA, SUNALT, MOONALT, MOONSEP, MOONFRAC
  • Trained on observations with moon altitude > 5°, moon fraction > 0.5

• ldm_twilight: Twilight sky spectra

  • 4 conditioning features: OBSALT, TRANSPARENCY_GFA, SUNALT, SUNSEP
  • Trained on observations with sun altitude > -20° (twilight conditions)

Improvements

• ✅ Comprehensive documentation for all three LDM variants
• ✅ Updated inference notebook with model comparison examples
• ✅ Complete training guides for all model types
• ✅ Integration tests for model loading and inference
• ✅ Automatic download from HuggingFace with SHA256 verification

Installation

pip install --upgrade desisky[cpu,data]

Quick Start

from desisky.io import load_builtin
from desisky.inference import LatentDiffusionSampler

# Load models
vae, _ = load_builtin("vae")
ldm_twilight, _ = load_builtin("ldm_twilight")

# Create sampler and generate spectra
sampler = LatentDiffusionSampler(ldm_twilight, vae, method="heun")

See CHANGELOG.md for full details.

Closes: #10, #11