README.md

metadata

license: cc-by-nc-sa-4.0
task_categories:
  - image-to-image
tags:
  - face-restoration
  - reference-based-restoration
  - training-data
  - snapcv
pretty_name: RefineFIR TrainingData

RefineFIR TrainingData

This repository contains training artifacts for the RefineFIR reimplementation:

• celebref_list_256_jw_0403.pkl: same-identity (source_id, target_id) training pairs.
• celebaref_warped_snapCV_jw_0403.zip: precomputed Snap/XCV warped reference images for the above pairs.

The clean image root used by train.py is expected to be arranged as CelebHQRefForRelease/<identity>/<image>.png, where pair ids such as 00001_2 resolve to CelebHQRefForRelease/00001/2.png.

The historical training pipeline generated these warped references with Snap/XCV. The public inference code uses MediaPipe, but training can consume the precomputed warped images directly without releasing Snap/XCV code.