README.md

---
pretty_name: "FreeStyle Dataset: cref/sref LoRA Triplets"
task_categories:
- image-to-image
tags:
- image-to-image
- style-reference
- content-reference
- lora
- triplet
- prompt-provenance
size_categories:
- 100K<n<1M
---

# FreeStyle Dataset: cref/sref LoRA Triplets

This repository contains a file-based image triplet dataset for content-reference (`cref`) and style-reference (`sref`) training/evaluation. The exported payload is under [`cref_sref/`](./cref_sref/).

Each row in a `triplets.csv` file represents one training sequence with three images:

- **content**: image used as the content reference (`cref_0`)
- **style**: image used as the style reference (`sref_0`)
- **target**: image for the combined content+style result

The images are deduplicated within each source/role. A single image file can be reused by multiple triplet rows.

## Quick Start

Install lightweight dependencies:

```bash
pip install huggingface_hub pandas pillow
```

Download and inspect the metadata for one source without downloading all images:

```python
from huggingface_hub import hf_hub_download
import pandas as pd

repo_id = "Blue2Giant/FreeStyle_Dataset"
source = "qwen"  # one of: qwen, flux, illustrious

triplets_csv = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename=f"cref_sref/{source}/triplets.csv",
)
triplets = pd.read_csv(triplets_csv)

print(len(triplets))
print(triplets[[
    "sequence_id",
    "content_image_path",
    "style_image_path",
    "target_image_path",
    "content_generation_prompt",
    "style_generation_prompt",
    "target_generation_prompt",
]].head())
```

Download the three images for one triplet:

```python
from huggingface_hub import hf_hub_download
from PIL import Image
import pandas as pd

repo_id = "Blue2Giant/FreeStyle_Dataset"
source = "qwen"

triplets_csv = hf_hub_download(repo_id=repo_id, repo_type="dataset", filename=f"cref_sref/{source}/triplets.csv")
triplets = pd.read_csv(triplets_csv)
row = triplets.iloc[0]

paths = {}
for role in ["content", "style", "target"]:
    rel = row[f"{role}_image_path"]  # relative to cref_sref/<source>/
    paths[role] = hf_hub_download(
        repo_id=repo_id,
        repo_type="dataset",
        filename=f"cref_sref/{source}/{rel}",
    )

images = {role: Image.open(path).convert("RGB") for role, path in paths.items()}
print(row["sequence_id"], images["content"].size, images["style"].size, images["target"].size)
```

## Recommended Download Patterns

This is a large file-tree dataset. Prefer selective downloads rather than cloning/downloading the entire repository at once.

Download all CSV/JSON/README metadata, but no images:

```python
from huggingface_hub import snapshot_download

snapshot_dir = snapshot_download(
    repo_id="Blue2Giant/FreeStyle_Dataset",
    repo_type="dataset",
    allow_patterns=[
        "README.md",
        "cref_sref/README.md",
        "cref_sref/*/README.md",
        "cref_sref/*/summary.json",
        "cref_sref/*/*.csv",
    ],
)
print(snapshot_dir)
```

Download one complete source, including images:

```python
from huggingface_hub import snapshot_download

source = "qwen"  # qwen is the smallest; flux and illustrious are much larger
snapshot_dir = snapshot_download(
    repo_id="Blue2Giant/FreeStyle_Dataset",
    repo_type="dataset",
    allow_patterns=["README.md", "cref_sref/README.md", f"cref_sref/{source}/**"],
)
print(snapshot_dir)
```

If you need the full dataset, use `snapshot_download(repo_type="dataset")`, but expect many files.

## Repository Layout

```text
<repo-root>/
  README.md                         # this guide
  cref_sref/
    README.md                       # payload-level guide
    HF_UPLOAD_CHECKLIST.md
    qwen/
      README.md
      summary.json
      triplets.csv
      content_images.csv
      style_images.csv
      target_images.csv
      images/
        content/...
        style/...
        target/...
    flux/
      ... same layout ...
    illustrious/
      ... same layout ...
```

Public upload notes:

- `_state/` export-resume folders are omitted.
- `logs/` are omitted.
- Paths stored in `triplets.csv` are relative to the source directory, e.g. relative to `cref_sref/qwen/`.

## Sources and Counts

| Source | Triplet rows | Unique content images | Unique style images | Unique target images |
|---|---:|---:|---:|---:|
| `qwen` | 33,582 | 90 | 424 | 5,162 |
| `flux` | 273,682 | 482 | 13,037 | 129,237 |
| `illustrious` | 172,589 | 8,045 | 1,581 | 65,952 |
| **Total** | **479,853** | **8,617** | **15,042** | **200,351** |

Total exported image files across source/role directories: **224,010**.

## File Semantics

### `triplets.csv`

Use `triplets.csv` as the main training/evaluation index. Important columns include:

- `sequence_id`: unique sequence id
- `base_model`: one of `qwen`, `flux`, `illustrious`
- `pair_key`: pair identifier
- `content_model_id`, `style_model_id`
- `content_image_path`, `style_image_path`, `target_image_path`
- `content_original_path`, `style_original_path`, `target_original_path`
- `content_match_status`, `style_match_status`, `target_match_status`
- `content_prompt_status`, `style_prompt_status`, `target_prompt_status`
- `content_generation_prompt`, `style_generation_prompt`, `target_generation_prompt`
- `vault_texts_json`

For model training, the most common fields are the three `*_image_path` columns and, when available, the three `*_generation_prompt` columns.

### `content_images.csv`, `style_images.csv`, `target_images.csv`

These files are deduplicated image-level metadata tables. Important columns include:

- `exported_image_path`: image path relative to the source directory
- `original_path`: best-effort recovered original generation image path
- `match_status`: original-path matching status
- `prompt_status`: prompt recovery status
- `generation_prompt`
- `base_prompt`
- `sequence_count`: number of triplet rows that reuse this image
- `sequence_ids_json`: triplet ids that reuse this image

Join keys:

- `triplets.csv.content_image_path` -> `content_images.csv.exported_image_path`
- `triplets.csv.style_image_path` -> `style_images.csv.exported_image_path`
- `triplets.csv.target_image_path` -> `target_images.csv.exported_image_path`

Example join:

```python
from huggingface_hub import hf_hub_download
import pandas as pd

repo_id = "Blue2Giant/FreeStyle_Dataset"
source = "qwen"

triplets = pd.read_csv(hf_hub_download(repo_id=repo_id, repo_type="dataset", filename=f"cref_sref/{source}/triplets.csv"))
content_meta = pd.read_csv(hf_hub_download(repo_id=repo_id, repo_type="dataset", filename=f"cref_sref/{source}/content_images.csv"))
content_meta = content_meta.set_index("exported_image_path")

row = triplets.iloc[0]
print(content_meta.loc[row["content_image_path"]])
```

## Match and Prompt Status Values

`match_status` describes whether the exporter could map an exported vault image back to an original candidate image:

- `matched`: exact visual-key match found in the candidate pool
- `unmatched`: candidate pool existed, but no exact unique match was found
- `ambiguous`: more than one candidate matched the same visual key
- `no_candidates`: no candidate pool was available for that lookup

`prompt_status` describes whether generation prompt metadata was recovered:

- `resolved`: prompt metadata was recovered
- `unmatched_original`: original image path was not matched
- `missing_prompt_payload`: prompt sidecar JSON was missing
- `missing_prompt_entry`: prompt file existed, but the specific image entry was missing
- `missing_prompt_index`: image filename could not be mapped to a prompt index

## Important Notes

- Exported images are vault training images, not guaranteed to be raw copies of the original one-LoRA or dual-LoRA generation files.
- `original_path` and prompt recovery fields are best-effort provenance fields. Do not assume every row has a resolved prompt.
- Rows with unresolved provenance are intentionally kept rather than forcing incorrect prompt assignments.
- The repository is organized as ordinary files. Use `huggingface_hub`, `pandas`, and `Pillow` for flexible access.

## Where to Start

1. Choose a source: start with `qwen` for a smaller first pass.
2. Read `cref_sref/<source>/README.md` and `summary.json`.
3. Load `cref_sref/<source>/triplets.csv`.
4. For each row, resolve image paths relative to `cref_sref/<source>/`.
5. Optionally join to the deduplicated `*_images.csv` metadata tables for provenance and reuse statistics.