tjw/deepfake-celeba

MetaFLOS Deepfake Dataset (CelebA Face Deepfake Generated Images)

Flux2-Klein generated deepfake images from CelebA face descriptions, for deepfake detection / comparison research.

Content

• 19,867 generated images (full CelebA validation split)
• Resolution: 256×256
• Each corresponds to a CelebA real face (see real_orig field in manifest)
• Generation style: snapshot/crop candid-photo look (not studio portrait) — off-center framing, subject possibly touching or cut by the frame edge, realistic film grain, softened contrast, natural non-HDR exposure

⚠️ Important: real images are NOT included

This dataset contains only generated images (generated/), not the CelebA real face images (privacy & licensing considerations).

The correspondence is recorded in the real_orig field of manifest.jsonl, e.g.:

"real_orig": "images/validation/faces/000000.jpg"

This is a path inside the original CelebA dataset. To obtain the corresponding real image, download the CelebA dataset yourself (img_align_celeba or fixed768_celeba_full_structured_prompts version) and use the real_orig path.

Directory Structure

deepfake_celeba/
├── generated/      19,867 fake images (256×256)
│   ├── 000/ ... 019/   20 subdirectories (≤1,000 images each; HF limit is 10,000 files/dir)
└── manifest.jsonl  19,867 records (with real_orig correspondence)

Manifest Schema (JSONL, one record per line)

{
  "sample_key": "validation:faces:000000",
  "real_orig": "images/validation/faces/000000.jpg",   // ← CelebA original path (corresponding real)
  "real_orig_size": [768, 768],
  "real": "real/validation_faces_000000.png",           // local thumbnail path (not uploaded, reference only)
  "real_size": [256, 256],
  "fake": "generated/000/validation_faces_000000.png",
  "fake_size": [256, 256],
  "comparison": null,
  "prompt": "original gemma4 description prompt",
  "prompt_used": "actual prompt with snapshot/crop style prefix",
  "seed": 9204520440607619239,
  "steps": 4,
  "cfg": 1.0
}

Generation Method

• fake: ComfyUI Flux2-Klein (flux-2-klein-9b-kv-fp8 + qwen_3_8b_fp8mixed) pure text-to-image, steps=4, cfg=1, seed deterministically derived from sample_key
• prompt: gemma4 detailed face description + snapshot/crop style prefix (off-center crop feel, softened contrast, non-HDR)
• prompt privacy: all prompts describe people with generic attributes (age, gender, ethnicity, hair, expression, clothing), no personal names (programmatically verified)
• Generation script: scripts/gen_celeba_deepfake.py (deterministic seed → reproducible)

Privacy Protection

• No real personal images: generated/ contains only model-synthesized images, not photos of any real person.
• No personal names in prompts: people are described with generic attributes; all 19,867 records × 2 fields (prompt/prompt_used) verified to contain no names.
• No EXIF/GPS metadata: generated images are pure PNG pixel data, with no location/device/time metadata.
• Real images not distributed: only the real_orig path is recorded; users must obtain real images per the CelebA license.

License & Copyright

• This dataset (generated images + manifest): MetaFLOS Research Use License v1.0 — see the LICENSE file for the binding terms. Academic research and non-commercial use only.
• Generated images: all synthesized by FLUX.2 [klein] 9B; no real personal images, no portrait rights involved.
• ⚠️ Why non-commercial: two independent reasons. (1) The corresponding real images belong to CelebA (non-commercial research purposes only); (2) synthetic facial imagery carries misuse risk, so binding safeguards are required (see Synthetic-content labeling and No re-identification in the LICENSE). Note on the generation model: FLUX.2-klein-9B is distributed under the FLUX Non-Commercial License v2.1 and is subject to the BFL Acceptable Use Policy; under its §2(d) the FLUX license does not itself prohibit commercial use of Outputs, but it does require content filtering or pre-distribution review and AI-generated disclosure (§2(e)), and prohibits using Output for surveillance, biometric processing, or to train models competitive with FLUX (§2(d), §4(a)) — obligations dataset users should observe as well.
• Corresponding real images: belong to the CelebA dataset (non-commercial research purposes only), copyright held by the original CelebA authors; not redistributed here — obtain them yourself per the CelebA license.
• Other generation components: Qwen3-8B text encoder (Apache 2.0); prompt field face descriptions produced by a Gemma-family model, generated in compliance with the Gemma Prohibited Use Policy (prompts are Gemma Outputs, which Gemma Terms §1.1(e)/§3.3 exclude from the §3.1 redistribution conditions — but a model fine-tuned on them may itself become a Gemma Model Derivative).
• Synthetic labeling required: when displayed or redistributed, images must be clearly labeled as AI-generated (cf. EU AI Act transparency obligations).
• Prohibited uses: commercial use; identity impersonation; building facial-recognition, biometric, or personal-image databases; re-identification against CelebA or any face corpus; non-consensual intimate imagery; fraud, disinformation, or election interference; any use violating local law.

Author & Citation

Tzer-jen Wei

@misc{metaflos_deepfake_celeba_2026,
  title  = {MetaFLOS Deepfake Dataset (CelebA Face Deepfake Generated Images)},
  author = {Wei, Tzer-jen},
  year   = {2026},
  note   = {Generated with FLUX.2-klein-9B; prompts derived from CelebA descriptions},
  url    = {https://huggingface.co/datasets/tjw/deepfake-celeba}
}

---

中文說明

MetaFLOS Deepfake 資料集（CelebA 人臉 deepfake 生成圖）

Flux2-Klein 依 CelebA 人臉描述 prompt 生成的 deepfake 圖，用於 deepfake 偵測/對比研究。

內容

• 19,867 張生成圖（對應 CelebA validation split 完整）
• 解析度：256×256
• 每張對應一個 CelebA 真實人臉（見 manifest 的 real_orig 欄位）
• 生成風格：snapshot/crop 生活照裁切感（非棚拍肖像）——off-center 構圖、主體可能觸及/被畫面邊緣裁切、真實底片顆粒、柔化對比、自然非 HDR 曝光

⚠️ 重要：real 圖未包含在本資料集

本資料集只包含生成圖（generated/），不含 CelebA 真實人臉圖（隱私與授權考量）。

對應關係記錄在 manifest.jsonl 的 real_orig 欄位，例如：

"real_orig": "images/validation/faces/000000.jpg"

這是原始 CelebA 資料集內的路徑。若要取得對應的 real 圖，請自行下載 CelebA 資料集（img_align_celeba 或 fixed768_celeba_full_structured_prompts 版本），依 real_orig 路徑取用。

目錄結構

deepfake_celeba/
├── generated/      19,867 張 fake 圖（256×256）
│   ├── 000/ ... 019/   分 20 個子目錄（每目錄 ≤1,000 張，因 HF 每目錄限 10,000 檔）
└── manifest.jsonl  19,867 筆（含 real_orig 對應欄位）

Manifest Schema（JSONL，每行一筆）

{
  "sample_key": "validation:faces:000000",
  "real_orig": "images/validation/faces/000000.jpg",   // ← CelebA 原始路徑（對應的 real）
  "real_orig_size": [768, 768],
  "real": "real/validation_faces_000000.png",           // 本機縮圖路徑（未上傳，僅供參考）
  "real_size": [256, 256],
  "fake": "generated/000/validation_faces_000000.png",
  "fake_size": [256, 256],
  "comparison": null,
  "prompt": "原始 gemma4 描述 prompt",
  "prompt_used": "加上 snapshot/crop 風格前綴的實際 prompt",
  "seed": 9204520440607619239,
  "steps": 4,
  "cfg": 1.0
}

生成方式

• fake：ComfyUI Flux2-Klein（flux-2-klein-9b-kv-fp8 + qwen_3_8b_fp8mixed）純 text-to-image，steps=4、cfg=1、seed 由 sample_key 決定性產生
• prompt：gemma4 產的詳細人臉描述 + snapshot/crop 風格前綴（off-center 裁切感、柔化對比、非 HDR）
• prompt 隱私：所有 prompt 皆以泛化描述（年齡、性別、種族、髮型、表情、服裝）描述人物，不含任何個人姓名（已程式掃描確認）
• 生成腳本：scripts/gen_celeba_deepfake.py（seed 決定性 → 可重現）

隱私保護

• 無真實個人影像：generated/ 全部為模型合成圖，非任何真實人物的照片。
• prompt 無個人姓名：人物一律以泛化特徵描述，已掃描 19,867 筆 × 2 欄位（prompt/prompt_used）確認不含人名。
• 無 EXIF/GPS 元資料：生成圖為純 PNG 像素資料，不含拍攝位置、裝置、時間等隱私元資料。
• real 圖不隨資料集發布：僅以 real_orig 欄位記錄 CelebA 原始路徑，需使用者自行依 CelebA 授權取得，本資料集不代為發布任何真實人臉影像。

授權與版權

• 本資料集（生成圖 + manifest）：MetaFLOS Research Use License v1.0，完整條款見 LICENSE 檔案。僅供學術研究與非商業用途。
• 生成圖：全部為 FLUX.2 [klein] 9B 合成，無真實個人影像，不涉及個人肖像權。
• ⚠️ 為何必須非商業：兩個獨立理由。(1) 對應的 real 圖屬 CelebA（僅供非商業研究用途）；(2) 合成人臉影像有濫用風險，需有拘束力之保護條款（見 LICENSE 之「合成內容標示」與「禁止再識別」）。關於生成模型：FLUX.2-klein-9B 採 FLUX 非商業授權 v2.1 並受 BFL Acceptable Use Policy 規範；依其 §2(d)，FLUX 授權本身不禁止 Output 之商業使用，但課予內容過濾或散布前審查、AI 生成標示之義務（§2(e)），並禁止以 Output 用於監控、生物特徵處理或訓練與 FLUX 競爭之模型（§2(d)、§4(a)）——資料集使用人亦應一併遵守。
• 對應的 real 圖：屬 CelebA 資料集（僅供非商業研究用途），版權歸 CelebA 原始作者所有；本資料集未再散布，請依 CelebA 授權自行取得。
• 其他生成元件：Qwen3-8B 文字編碼器（Apache 2.0）；prompt 欄位之人臉描述由 Gemma 系列模型產生，已依 Gemma 禁止用途政策生成（prompt 屬 Gemma Outputs，依 Gemma 條款 §1.1(e)/§3.3 不受 §3.1 散布條件拘束 —— 惟以其微調而得之模型本身可能構成 Gemma Model Derivative）。
• 須標示合成內容：展示或再散布時，須清楚標示影像為 AI 生成（參照歐盟 AI Act 透明性義務）。
• 禁止用途：商業使用；身分冒充；建立人臉辨識、生物特徵或個人影像資料庫；與 CelebA 或任何人臉語料進行再識別比對；未經同意之私密影像；詐欺、不實資訊或干預選舉；任何違反當地法律之用途。

著作人與引用

Tzer-jen Wei

引用格式同英文段落之 BibTeX。