| --- |
| pretty_name: Real Downscale Upscaling Restoration Dataset |
| task_categories: |
| - image-to-image |
| language: |
| - en |
| size_categories: |
| - 1K<n<10K |
| configs: |
| - config_name: default |
| drop_labels: true |
| --- |
| |
| # Real Downscale Upscaling Restoration Dataset |
|
|
| This dataset contains 9734 reproducibly sampled anime images for image |
| super-resolution and detail restoration. |
|
|
| Each row contains: |
|
|
| - `ref1`: a low-resolution full-image reference generated by downscaling the |
| original image while preserving aspect ratio. |
| - `target`: the original clean full-resolution image, saved as lossless WebP by |
| default. |
| - `edit_instruction`: an English instruction asking the model to upscale the |
| image and restore missing details. |
| - metadata columns: source paths, captions, descriptions, downscale ratio, |
| ref1 dimensions, original dimensions, and seeds. |
|
|
| The dataset was generated with seed `251101`. |
|
|
| Downscale ratios used for ref1: `0.75, 0.5, 0.25, 0.1`. |
|
|
| ## Hugging Face layout |
|
|
| ```text |
| train/ |
| ref1/ |
| prompt/ |
| target/ |
| metadata.jsonl |
| README.md |
| ``` |
|
|
| `metadata.jsonl` uses `ref1_file_name` and `target_file_name` so Hugging Face |
| ImageFolder can display two image columns in the Dataset Viewer. |
|
|
| This dataset is generated by CPU image resizing only. It does not use model |
| inference. WebP is used by default to reduce storage while keeping Hugging Face |
| Dataset Viewer compatibility. |
|
|
