🌙 Lyra/Lune Flow-Matching Image Generation

Geometric crystalline diffusion with flow matching by AbstractPhil

Generate images using SD1.5-based models with geometric deep learning:

  • Flow-Lune: Flow matching with pentachoron geometric structures (15-25 steps)
  • SD1.5 Base: Standard Stable Diffusion 1.5 baseline
  • Lyra VAE Toggle: Add CLIP+T5 fusion for side-by-side comparison
  • CLIP Variants: Different text encoders for varied semantic understanding

Enable Lyra to see both standard CLIP and geometric CLIP+T5 fusion results!
Base Model
CLIP Model

Text encoder variant

Generate side-by-side comparison with geometric fusion

Use flow matching ODE integration
0 5
Prediction Type

Type of model prediction
1 50
1 20
256 1024
256 1024
0 4294967295

Tips:

  • Flow matching works best with 15-25 steps (vs 50+ for standard diffusion)
  • Shift controls the flow trajectory (2.0-2.5 recommended for Lune)
  • Lower shift = more direct path, higher shift = more exploration
  • Lune uses v_prediction by default for optimal results
  • Lyra toggle generates side-by-side comparison (CLIP vs CLIP+T5 fusion)
  • CLIP variants may give different semantic interpretations
  • SD1.5 Base uses epsilon (standard diffusion)
  • Lune operates in a scaled latent space (5.52x) for geometric efficiency

Model Info:

  • Flow-Lune: Trained with flow matching on 500k SD1.5 distillation pairs
  • Lyra VAE: Multi-modal fusion (CLIP+T5) via Cantor geometric attention
  • SD1.5 Base: Standard Stable Diffusion 1.5 for comparison

CLIP Models:

  • openai/clip-vit-large-patch14: Standard CLIP-L (default)
  • openai/clip-vit-large-patch14-336: Higher resolution CLIP-L
  • laion/CLIP-ViT-L-14: LAION-trained CLIP-L variant
  • laion/CLIP-ViT-bigG-14: Larger CLIP-G model

📚 Learn more about geometric deep learning
Examples
Prompt Negative Prompt Base Model CLIP Model Steps CFG Scale Width Height Shift Enable Flow Matching Prediction Type Enable Lyra VAE (CLIP+T5 Fusion) Seed Randomize Seed