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.