image image |
|---|
IDCFIQA: A Dual-Dimension Quality Assessment Benchmark Dataset for Identity-Consistent Face Image Generation
Overview
Identity-consistent (ID-consistent) human face image generation has become a core task in virtual reality, film production, digital avatars, and personalized entertainment. Despite the recent proliferation of such generative methods, the absence of specialized quality assessment benchmarks hampers fair algorithmic comparisons, ultimately constraining the advancement of ID-consistent face image generation techniques.
To bridge this gap, we propose IDCFIQA, an ID-consistent face image generation quality assessment benchmark dataset. IDCFIQA consists of 200 source face images and 1,600 generated face images produced by eight representative generation algorithms. The source images include 100 real face images and 100 synthetic face images.
To provide a comprehensive evaluation, we conduct subjective pairwise annotations and employ the Bradley-Terry (B-T) model to derive scores across two essential dimensions:
- Perceptual Quality: the visual quality of the generated face images.
- ID Consistency: the identity fidelity between the generated images and the corresponding source images.
Leveraging this benchmark, we evaluate a wide range of existing image quality assessment methods and provide an in-depth analysis of their capability to accurately assess both visual quality and identity consistency.
Dataset Summary
| Item | Description |
|---|---|
| Dataset Name | IDCFIQA |
| Task | Quality assessment for identity-consistent face image generation |
| Total Images | 1,800 |
| Source Images | 200 |
| Real Source Images | 100 images from AgeDB |
| Synthetic Source Images | 100 images generated by ThisPersonDoesNotExist |
| Generated Images | 1,600 |
| Generation Methods | 8 representative ID-consistent face generation algorithms |
| Evaluation Dimensions | Perceptual quality and ID consistency |
| Annotation Type | Human pairwise comparison |
| Score Estimation | Bradley-Terry (B-T) model |
| Human Votes | 84,000 valid pairwise votes |
Usage
You can download the IDCFIQA dataset and code from the following links:
BaiduYun Disk
Link: https://pan.baidu.com/s/1COe-R4kTm6C1-V7aq3kGvw?pwd=dd8j
Password: dd8j
Google Drive
Link: https://drive.google.com/drive/folders/1YOv7jCwPeyZPtT63EXjw2NK-m6-RqHWH?usp=sharing
Minimal Responsible AI Metadata
Data Limitations
RAI field: rai:dataLimitations
IDCFIQA is specifically designed for quality assessment of identity-consistent face image generation. Its applicability is mainly limited to evaluating generated human face images under two dimensions: perceptual visual quality and identity consistency.
The dataset contains 200 source face images, including 100 real images from AgeDB and 100 synthetic identities generated by ThisPersonDoesNotExist, together with 1,600 generated images produced by eight representative ID-consistent face generation methods.
Therefore, the dataset may not fully cover all demographic groups, age ranges, facial attributes, lighting conditions, poses, occlusions, cultural contexts, or real-world deployment scenarios.
The dataset is not recommended for:
- General face recognition training
- Biometric identification
- Surveillance
- Demographic classification
- Security-sensitive identity verification
- Evaluation of non-face image generation models
Data Biases
RAI field: rai:dataBiases
Potential biases may arise from source image selection, the limited number of source identities, the demographic composition of annotators, and the choice of generation algorithms.
Although the source images were curated to include diverse ethnicities and skin tones, the dataset contains only 200 source identities. Therefore, some demographic groups, age ranges, facial appearances, and imaging conditions may be underrepresented.
The subjective annotations were collected from 150 undergraduate and graduate students aged 18–30 with normal or corrected-to-normal visual acuity, which may introduce annotator demographic bias.
In addition, using a single prompt, “a photo of a person,” reduces prompt-related variability but may limit the diversity of semantic generation conditions. Algorithmic bias may also be introduced by the eight selected face generation methods, whose performance differs between real and synthetic subsets.
Personal or Sensitive Information
RAI field: rai:personalSensitiveInformation
Yes. The dataset contains human face images, which may be considered biometric and personally sensitive information.
The real subset is manually selected from the public AgeDB database, while the synthetic source subset is generated using ThisPersonDoesNotExist.
The images may implicitly reveal or encode sensitive attributes such as:
- Apparent age
- Gender presentation
- Skin tone
- Ethnicity-related appearance
- Other facial characteristics
The current paper does not describe collecting explicit demographic labels, medical information, political or religious beliefs, socioeconomic status, or other non-visual personal metadata.
The dataset should be released and used only under research-oriented restrictions, with clear limitations against surveillance, biometric identification, or malicious identity manipulation.
Data Use Cases
RAI field: rai:dataUseCases
The dataset is intended to measure two real-world constructs:
- Perceptual visual quality of generated face images
- Identity consistency between generated and source faces
Construct validity is supported through human pairwise preference annotations and Bradley-Terry score estimation across these two dimensions.
Validated use cases include:
- Benchmarking image quality assessment methods for ID-consistent face generation
- Comparing identity-preserving face generation algorithms
- Analyzing the trade-off between visual quality and identity fidelity
- Developing new IQA metrics for generated face images
Validity has not been established for:
- Face recognition
- Biometric verification
- Demographic inference
- Fairness auditing across protected groups
- Surveillance
- Real-world identity matching
- General non-face image quality assessment
Social Impact
RAI field: rai:dataSocialImpact
IDCFIQA may have positive societal impact by supporting more reliable and human-aligned evaluation of identity-consistent face generation methods. It can help researchers identify identity drift, generative artifacts, unnatural textures, over-smoothing, and limitations of existing IQA models.
However, because the dataset involves face generation and identity preservation, it may also carry risks of misuse, including improving face synthesis systems for impersonation, deepfake generation, or identity manipulation.
Fairness concerns may arise if the dataset underrepresents certain demographic groups or if evaluation models trained or selected using this dataset perform unevenly across populations.
Suggested mitigations include:
- Restricting the dataset to research and evaluation
- Documenting demographic and source limitations
- Prohibiting use in surveillance or biometric identification
- Encouraging subgroup analysis when demographic annotations or ethically appropriate proxies are available
Synthetic Data
RAI field: rai:hasSyntheticData
True.
The dataset contains synthetic data in two forms:
- 100 source identity images are generated using ThisPersonDoesNotExist.
- 1,600 generated face images are produced by eight identity-consistent face generation algorithms.
The generation process uses a standardized prompt:
a photo of a person
Source Datasets
RAI field: prov:wasDerivedFrom
The real source images are derived from the public AgeDB dataset.
The synthetic source images are derived from ThisPersonDoesNotExist.
The generated images are further derived from eight ID-consistent face generation methods:
- IPA-FaceID-Plus
- IPA-FaceID-Plus V2
- Arc2Face
- FastComposer
- InstantID
- IPA-FaceID (SDXL)
- PhotoMaker
- AlvPai
Official URIs for AgeDB, ThisPersonDoesNotExist, and the corresponding generation method repositories or webpages should be provided where available.
Provenance Activities
RAI field: prov:wasGeneratedBy
Data Collection
The dataset begins with 200 high-quality source face images, including 100 real-world images manually selected from AgeDB and 100 synthetic identity images generated by ThisPersonDoesNotExist.
The images are selected for high resolution and curated to cover diverse ethnicities and skin tones.
Data Generation
Eight representative ID-consistent face generation methods are applied to each source image using the same prompt:
a photo of a person
This process produces 1,600 generated face images. Together with the 200 source images, the benchmark contains 1,800 images in total.
Data Annotation
The subjective user study follows a pairwise comparison protocol. Participants are shown two generated face images produced by different algorithms for the same source image and are asked to select the better one.
The paper reports:
- 150 undergraduate and graduate student participants
- Participants aged 18–30
- Normal or corrected-to-normal visual acuity
- 84,000 valid human votes
Bradley-Terry modeling is used to convert pairwise comparisons into subjective scores for perceptual quality and identity consistency.
Preprocessing and Quality Control
The subjective experiments follow ITU-R BT.500-14 recommendations and use controlled display conditions.
- Downloads last month
- 926