Mortallll/Syn-UI-Final-samples

Syn-UI-Final Method Samples

This repository contains a small visual preview of the first four Syn-UI-Final generation methods.

It includes 200 generated images:

• syreanet: 50 images
• watergan: 50 images
• uwnr: 50 images
• cut: 50 images

For each method, 25 images were sampled from the training split and 25 images from the test split using a fixed seed (20260629). The sample is intended for quick visual inspection only; it is not the full Syn-UI-Final dataset.

Directory structure

images/
  syreanet/
    train/
    test/
  watergan/
    train/
    test/
  uwnr/
    train/
    test/
  cut/
    train/
    test/
metadata.csv

Methods

• SyreaNet: depth-aware physical/rule-based underwater synthesis.
• WaterGAN: official WaterGAN training flow with depth and Real-U water-domain images. The preview uses the corrected TensorFlow compatibility environment and the selected official-training checkpoint DCGAN.model-1442.
• UWNR: pretrained network-based underwater image generation using clean images, pseudo-depth, and water reference information.
• CUT: unpaired image-to-image translation from ImageNet-style natural images to Real-U-style underwater images.

Notes

• Only generated preview images are included here.
• The 50 WaterGAN previews were refreshed from the final watergan_official_corrected_v2 dataset after full count, readability, size, and sampling-policy validation.
• Original server paths, account credentials, and full dataset manifests are intentionally not included.
• This preview is for research communication and qualitative comparison. Please check the licenses and usage terms of the original data sources before redistributing or using a full derived dataset.