README.md

---
pretty_name: Anime Background Edit Pairs (WebDataset)
task_categories:
- image-to-image
language:
- en
size_categories:
- 10K<n<100K
tags:
- anime
- image-editing
- paired-images
- webdataset
configs:
- config_name: default
  data_files:
  - split: train
    path: train-*.tar
---

# Anime Background Edit Pairs (WebDataset)

This repository is the canonical distribution of **20,221 paired anime
background edits**, packaged as 10 streaming-friendly WebDataset TAR shards.
It is self-contained and does not require a separate ImageFolder repository.

Each sample contains a source image (`ref1`), an English editing instruction,
and the desired edited image (`target`). Source and target images come from the
same background group, so the instruction describes the visual transformation
while preserving the underlying scene composition.

## Dataset summary

| Property | Value |
|---|---:|
| Samples | 20,221 |
| TAR shards | 10 |
| TAR members | 80,884 |
| Total shard size | 7,118,909,440 bytes (about 6.63 GiB) |
| Image format | WebP, quality 95 |
| Instruction language | English |
| Invalid-pair flags retained | 183 |
| Abandoned pairs excluded | 9 |

All TAR members are regular files. Hardlinks used by the local source dataset
were dereferenced when the shards were created, making the archives safe for
sequential WebDataset streaming.

## Sample structure

Every sample uses a six-digit key and has four members:

```text
000000.ref1.webp
000000.target.webp
000000.prompt.txt
000000.json
```

- `ref1.webp`: source image to edit.
- `target.webp`: desired edited image.
- `prompt.txt`: executable English editing instruction.
- `json`: complete sample metadata and provenance.

The JSON member includes the following main fields:

- `id`, `sample_id`, `pair_id`, and `group_id`
- `edit_instruction`
- `difference_summary`, `changes`, and `preserve`
- `annotation_model` and `annotation_source`
- `confidence`, `invalid_pair`, and `review_status`
- source/target dimensions and SHA-256 hashes
- `ref1_member`, `target_member`, and `prompt_member`

## Dataset creation

The source illustrations were organized into background groups containing
multiple visual variants of the same scene. Images within each group were paired
to form source/target edit examples. A vision-language model compared both
images and produced a structured English instruction describing the changes to
apply while preserving the scene's composition and unaffected objects.

Images were converted to WebP at quality 95 while retaining their original
dimensions. The completed records were assigned stable keys from `000000` to
`020220`, then packed into 10 deterministic shards. Each image is stored as a
regular TAR member, including samples that reuse the same underlying source
image, so the dataset can be consumed safely by sequential streaming readers.

## Visual examples

Each row below is one complete training example: the model receives the source
image and instruction, and learns to produce the target image.

### Daytime to nighttime (`000073`, group `BG000003`)

**Instruction:** Change the scene from daytime to nighttime, applying a
blue-tinted darkening effect to the entire view.

| Source (`ref1`) | Target |
|---|---|
| ![Daytime residential street](assets/examples/000073_ref1.webp) | ![The same residential street at night](assets/examples/000073_target.webp) |

### Green landscape to autumn (`001492`, group `BG000292`)

**Instruction:** Change the landscape colors to an autumnal palette, replacing
green foliage with warm orange and red tones in the trees and grass.

| Source (`ref1`) | Target |
|---|---|
| ![Green landscape](assets/examples/001492_ref1.webp) | ![The same landscape in an autumn palette](assets/examples/001492_target.webp) |

## Loading with WebDataset

Because this repository is manually gated, first request access on the dataset
page and authenticate with Hugging Face:

```bash
hf auth login
pip install webdataset huggingface_hub pillow
```

Then download or reuse the authenticated local Hub cache and stream the shards:

```python
from pathlib import Path

import webdataset as wds
from huggingface_hub import snapshot_download

repo_dir = snapshot_download(
    repo_id="LAXMAYDAY/anime_background_edit_data_webdataset",
    repo_type="dataset",
    allow_patterns="train-*.tar",
)
shards = sorted(str(path) for path in Path(repo_dir).glob("train-*.tar"))

dataset = (
    wds.WebDataset(shards, shardshuffle=True)
    .shuffle(1000)
    .decode("pil")
    .to_tuple("ref1.webp", "target.webp", "prompt.txt", "json")
)

for ref1, target, instruction, metadata in dataset:
    print(instruction)
    print(metadata["pair_id"], metadata["group_id"])
    break
```

For distributed training, split shards by worker/node using the normal
WebDataset pipeline utilities appropriate to your training framework.

## Annotation provenance

- `doubao-seed-2-0-mini-260428`: 18,131 samples
- `gpt-5.5`: 2,090 samples

Annotation sources:

- `doubao_initial`: 15,699
- `doubao_final`: 2,432
- `jarless_middle`: 2,090

## Quality notes

The dataset retains 183 completed samples marked `invalid_pair=true`. These may
represent unrelated, unusable, or effectively unchanged pairs. They are kept so
downstream users can decide whether identity/no-op examples are useful; filter
them through the JSON metadata when needed.

Nine pairs that did not pass annotation validation are excluded from all
shards. Multiple edit pairs may reuse the same underlying image when one
background group contains more than two visual variants.

## Shards

```text
train-00000-of-00010.tar
train-00001-of-00010.tar
train-00002-of-00010.tar
train-00003-of-00010.tar
train-00004-of-00010.tar
train-00005-of-00010.tar
train-00006-of-00010.tar
train-00007-of-00010.tar
train-00008-of-00010.tar
train-00009-of-00010.tar
```

The first shard contains 2,023 samples; each remaining shard contains 2,022.