Initial release of Shadow Ink Sketch.
Trained on 50 synthetic images spanning people, animals, nature, objects, vehicles, and architecture.
Training settings:
• Steps: 1,000
• Effective dataset passes: 20
• Resolution: 512-pixel buckets
• Rank / alpha: 16 / 16
• Optimizer: AdamW
• Learning rate: 0.0001
• Scheduler: constant
• Batch size: 1
• Seed: 42
• Training time: 52 minutes 32 seconds
Training provenance:
The source images were generated with FLUX 3 Klein through WanGP. Training used Krea 2 RAW weights reconstructed from an installed INT8 checkpoint; quantization loss remains. The trained LoRA is for the Krea 2 architecture.
Inference quality and compatibility with specific workflows have not yet been validated.