---
license: unknown
task_categories:
- image-to-image
tags:
- super-resolution
- LSDIR
pretty_name: SeeSR ODCLIP LSDIR (aug)
---

# SeeSR / ODCLIP — LSDIR_aug training data

Augmented LSDIR training data used to train ODCLIP (SeeSR-style SD2+ControlNet real-world SR).
Packed as `.tar` shards (~5000 files each) to keep the file count HF-friendly.

## Contents (7 folders, 84,990 paired samples each; paired by numeric filename ID)
| folder | ext | content |
|---|---|---|
| `gt` | .png | HR ground-truth |
| `lr` | .png | degraded LR |
| `lr_bicubic` | .png | bicubic-upsampled LR |
| `latent_hr` | .pt | VAE latent of HR |
| `pool_embeds` | .pt | pooled text/tag embedding |
| `prompt_embeds` | .pt | prompt text embedding |
| `prompt_txt` | .txt | tag/prompt text |

Each folder is split into 17 shards: `<folder>_000.tar` … `<folder>_016.tar`.

## Extract
```bash
for t in *.tar; do tar -xf "$t"; done
```
This recreates `gt/`, `lr/`, `latent_hr/`, … with the original files.