FakeParts: A New Family of AI-Generated DeepFakes

Abstract

We introduce FakeParts, a new class of deepfakes characterized by subtle, localized manipulations to specific spatial regions or temporal segments of otherwise authentic videos. Unlike fully synthetic content, these partial manipulations—ranging from altered facial expressions to object substitutions and background modifications—blend seamlessly with real elements, making them particularly deceptive and difficult to detect. To address the critical gap in detection, we present FakePartsBench, the first large-scale benchmark specifically designed to capture the full spectrum of partial deepfakes. Comprising over 81K (including 44K FakeParts) videos with pixel- and frame-level manipulation annotations, our dataset enables comprehensive evaluation of detection methods. Our user studies demonstrate that FakeParts reduces human detection accuracy by up to 26% compared to traditional deepfakes, with similar performance degradation observed in state-of-the-art detection models. This work identifies an urgent vulnerability in current detectors and provides the necessary resources to develop methods robust to partial manipulations.

Summary

• Problem. Existing detectors and benchmarks mainly target fully synthetic videos, leaving subtle, localized edits (FakeParts) under-represented and highly deceptive in real scenarios.
• Solution. We formally define FakeParts as a new class of deepfakes with partial spatial or temporal manipulations and release FakePartsBench: a large benchmark with pixel-level and frame-level annotations that spans both full deepfakes and diverse partial edits.
• Finding. FakePartsBench contains over 81,000 videos (including ~44K FakeParts and ~20K full deepfakes), and both humans and SOTA detectors struggle significantly more with partial manipulations compared to fully synthetic content. Detection performance for FakeParts can drop up to ~50 % relative to traditional deepfakes.
• Use. This benchmark supports the development and evaluation of detectors that not only identify whether a video is manipulated but also where and when those manipulations occur.

Dataset 💽

FakePartsBench provides:

• 81,000+ videos total: ~44,000 FakeParts, ~20,000 full deepfakes, and ~17,000 real clips
• High-resolution content (many samples up to 1080p), durations typically 5–14 s
• Annotations: fine-grained pixel masks, frame-level labels, and metadata (prompt, source frame, resolution, FPS)
• Manipulation categories:
  • Full deepfakes: text-to-video (T2V), image-to-video (I2V), text-to-image-to-video (TI2V) generations (e.g., via Sora, Veo2, Allegro AI)
  • Spatial FakeParts: face swap, inpainting, outpainting edits on select regions
  • Temporal FakeParts: frame interpolation and other short-span temporal edits
  • Style FakeParts: subtle style or colour shifts without structural changes

Every sample includes rich metadata and, for FakeParts, precise localization of manipulated pixels or frames to enable detailed training and evaluation.

Sample Usage 🚀

You can easily load the FakePartsBench dataset using the Hugging Face datasets library:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("hi-paris/FakeParts")

# Inspect the data
print(dataset)

Citations ✍️

If you use FakeParts please cite:

@misc{liu2025fakepartsnewfamilyaigenerated,
    title={FakeParts: a New Family of AI-Generated DeepFakes}, 
    author={Ziyi Liu and Firas Gabetni and Awais Hussain Sani and Xi Wang and Soobash Daiboo and Gaetan Brison and Gianni Franchi and Vicky Kalogeiton},
    year={2025},
    eprint={2508.21052},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2508.21052}, 
}

License & Responsible Use 🔨

• Code: see LICENSE (default: BSD-3-Clause unless noted otherwise in subfolders).
• Dataset: released for research and defensive purposes only.
  • Do not attempt to identify private individuals.
  • Do not use for generating disinformation or harassment.
  • Faceswap content uses celebrity imagery to avoid sensitive personal data.
• Please comply with third-party model/data licenses cited in the paper and baselines/.