---
license: other
task_categories:
- image-to-image
language:
- en
tags:
- image-editing
- instruction-guided-editing
- restoration
- webdataset
- tr-hash
pretty_name: Complexity Atlas Image Edits
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: train/*.tar
  - split: validation
    path: validation/*.tar
  - split: test
    path: test/*.tar
---

# Complexity Atlas Image Edits

Complexity Atlas Image Edits is an aligned instruction-guided image editing
dataset derived from the normalized public-domain Complexity Atlas image bank.
It contains 336,245 explicit source/instruction/target triplets at 256 x 256.

## Dataset structure

Each WebDataset record contains:

```text
<edit_id>.source.webp
<edit_id>.target.webp
<edit_id>.txt
<edit_id>.json
```

- `source.webp` is a deterministic degraded image;
- `target.webp` is the unchanged normalized public-domain artwork;
- `txt` is an original imperative restoration instruction;
- `json` records transformation parameters, hashes, provenance, source URL,
  upstream revision, source license, and split identity.

## Scale

| Property | Value |
|---|---:|
| Editing pairs | 336,245 |
| Train | 329,510 |
| Validation | 3,428 |
| Test | 3,307 |
| WebDataset shards | 136 |
| Stored bytes | 11,302,287,360 |
| Edit families | 13 |
| Exact instruction variants | 126 |
| Maximum exact instruction share | 1.56% |

## Editing families

The first release covers color restoration, deblurring, underexposure and
overexposure correction, contrast and saturation restoration, detail recovery,
compression restoration, warm and cool cast correction, orientation
correction, surface-damage restoration, and missing-region restoration.

These are verifiable restoration operations. The release does not claim to
contain semantic object insertion, removal, identity changes, or unconstrained
creative edits.

## Construction and split integrity

One deterministic edit is selected per source artwork. The original target's
split is preserved, so an artwork cannot cross train, validation, and test.
The complete audit reports:

- zero source-identity leakage;
- zero rejected records;
- complete source/instruction/target/metadata membership;
- a non-trivial pixel change for every source;
- unique edit identifiers;
- exact instruction repetition below 5%.

See `audit_report.json`, `build_manifest.json`, and `files_manifest.json` for
machine-readable evidence.

## Provenance and licensing

The normalized target images originate from
`Mitsua/art-museums-pd-440k` at pinned revision
`fba945da78b36262eb9272067197cc28d06cffbf`. Every retained record is marked
CC0 and retains its source URL and upstream metadata.

Generated source images are deterministic mechanical transformations of those
targets. This repository does not impose a new blanket copyright restriction
on the public-domain image assets. The builder code is available separately
under Apache-2.0. Users remain responsible for reviewing the per-record
provenance for their intended jurisdiction and use.

## Builder

The reproducible builder and tests are available at:

https://github.com/Complexity-ML/complexity-atlas-images