TRAINING_NOTES.md

Training Notes (9B)

Summary

• Base model: FLUX.2-klein-base-9B
• LoRA rank: 16
• Optimizer: AdamW8bit
• Learning rate: 3e-5
• Training steps (tentative best): 600

Dataset

• Source images were collected from open-license panorama/HDRI archives on the web.
• License policy was centered on permissive sources (CC0 / Public Domain), with part of the set including CC BY assets.
• After preprocessing and filtering, training was run on a curated ERP pair set of about 1000 samples.

Preprocess / Pair Generation

• Target format: ERP (equirectangular) 2:1
• Source panoramas were normalized to ERP and cleaned (invalid aspect/failed reads removed).
• Control images were generated by:
  • building a green ERP canvas,
  • sampling 1-3 reference patches,
  • projecting patches with a pinhole-based spherical mapping,
  • handling seam continuity with wrap-aware placement.
• The model is trained to fill green regions while preserving visible reference context.