README.md

---
pretty_name: AffectHuman-43K
language:
- en
license: other
task_categories:
- image-to-image
- text-to-image
- image-classification
- audio-classification
tags:
- affective-computing
- controllable-generation
- multimodal
- emotion
- identity-preservation
- human-generation
size_categories:
- 10K<n<100K
extra_gated_heading: "Request access to AffectHuman-43K"
extra_gated_description: "Access is limited to non-commercial academic research, benchmarking, and evaluation. Requests are reviewed manually by the dataset maintainers."
extra_gated_button_content: "Request research access"
extra_gated_prompt: "By requesting access, you agree to the AffectHuman-43K research-only license, responsible-use terms, and upstream dataset license obligations."
extra_gated_fields:
  "Full name": text
  "Institution or affiliation": text
  "Institutional email": text
  "Country": country
  "Research purpose":
    type: select
    options:
    - "Academic research"
    - "Non-commercial benchmarking"
    - "Teaching or education"
    - "Other"
  "Brief project description": text
  "I agree to use the dataset only for non-commercial academic research, benchmarking, or evaluation": checkbox
  "I agree not to use the dataset for biometric identification, surveillance, tracking, profiling, or face-recognition deployment": checkbox
  "I agree not to use the dataset for impersonation, deceptive media, harassment, or non-consensual synthetic media": checkbox
  "I agree not to redistribute, mirror, sell, sublicense, or publicly rehost the raw dataset files": checkbox
  "I agree to cite AffectHuman-43K and comply with all upstream dataset and model licenses": checkbox
---

# AffectHuman-43K

AffectHuman-43K is an emotion-aligned multimodal benchmark for controlled human affect generation and evaluation.

The benchmark contains 42,469 usable samples with complete image, reference-image, audio, and text coverage. Identity is specified through a visual reference image, while text, audio, and emotion labels provide affective control signals. This design separates identity preservation from affective control, enabling evaluation of whether a model can preserve a reference identity while following multimodal emotion-control inputs.

## Benchmark Purpose

AffectHuman-43K is designed as a controlled evaluation protocol rather than a naturally co-recorded multimodal identity dataset. The target task is:

```text
reference image -> identity condition
text/audio/emotion label -> affective control condition
target image -> supervision or evaluation target
```

This structure supports evaluation of:

- reference-conditioned identity preservation
- multimodal affective controllability
- repeated-reference identity consistency
- leakage-safe generalization to held-out reference identities

Reference-identity groups are defined by shared visual reference-image anchors. Samples sharing the same reference image are assigned to the same reference-identity group, and each group is placed wholly within a single split. This prevents cross-split reference-identity leakage.

## Access Model

The dataset page is public for transparency, citation, and reproducibility. The underlying dataset files are gated and require manual approval before download.

Access is intended for non-commercial academic research, benchmarking, and evaluation. The dataset contains human face/reference-image and speech-derived modalities, so users must agree to the responsible-use terms before access is granted.

Before approval, requesters must provide:

- full name
- institution or affiliation
- institutional email
- country
- research purpose
- brief project description
- agreement to non-commercial research-only use
- agreement not to use the dataset for biometric identification, surveillance, tracking, profiling, or face-recognition deployment
- agreement not to use the dataset for impersonation, deceptive media, harassment, or non-consensual synthetic media
- agreement not to redistribute or publicly mirror the raw files
- agreement to cite AffectHuman-43K and comply with upstream dataset and model licenses

## Dataset Summary

| Statistic | Value |
|---|---:|
| Total usable samples | 42,469 |
| Aligned candidate pool before leakage repair | 43,514 |
| Reference-identity groups | 26,241 |
| Repeated reference-identity groups | 10,204 |
| Singleton reference-identity groups | 16,037 |
| Multi-emotion repeated groups | 7,290 |
| Train / validation / test samples | 29,728 / 6,371 / 6,370 |
| Cross-split reference-identity leakage | 0 |
| Image coverage | 42,469 / 42,469 |
| Reference-image coverage | 42,469 / 42,469 |
| Audio coverage | 42,469 / 42,469 |
| Text coverage | 42,469 / 42,469 |

## Reference-Identity Grouping

A reference-identity group is defined by the shared identity reference image anchor. This means that all samples using the same reference image are treated as belonging to the same identity group for split construction and leakage checking.

This grouping protocol is used for two reasons:

- It prevents the same reference identity from appearing across train, validation, and test splits.
- It enables identity-consistency evaluation across multiple affective controls when a reference identity appears in more than one sample.

The benchmark contains both singleton and repeated reference-identity groups. Singleton groups support reference-conditioned unseen-identity evaluation, while repeated groups support consistency analysis under multiple affective transformations.

## Group Size Distribution

```text
1 sample per reference-identity group: 16,037 groups
2 samples per reference-identity group:  6,403 groups
3 samples per reference-identity group:  2,432 groups
4 samples per reference-identity group:    840 groups
5 samples per reference-identity group:    332 groups
6 samples per reference-identity group:    121 groups
7 samples per reference-identity group:     42 groups
8 samples per reference-identity group:     22 groups
9 samples per reference-identity group:      7 groups
10 samples per reference-identity group:     4 groups
11 samples per reference-identity group:     1 group
```

Most repeated reference-identity groups contain 2-5 samples, with a small number of larger groups.

## Emotion Classes

AffectHuman-43K uses an 8-class affective taxonomy. Emotion labels are unified across source datasets using valence-arousal alignment.

| Emotion Class | Samples |
|---|---:|
| Surprised | 7,272 |
| Neutral | 7,066 |
| Happy | 7,023 |
| Sad | 5,667 |
| Angry | 4,759 |
| Disgusted | 4,216 |
| Fearful | 3,404 |
| Contemptuous | 3,062 |
| Total | 42,469 |

## Split Distribution

| Split | Samples |
|---|---:|
| Train | 29,728 |
| Validation | 6,371 |
| Test | 6,370 |
| Total | 42,469 |

No reference-identity group appears in more than one split.

## Folder Structure

```text
AffectHuman-43K/
├── README.md
├── images/
│   ├── train/{emotion}/{sample_id}.jpg
│   ├── val/{emotion}/{sample_id}.jpg
│   └── test/{emotion}/{sample_id}.jpg
├── reference_images/
│   ├── train/{emotion}/{sample_id}_ref.jpg
│   ├── val/{emotion}/{sample_id}_ref.jpg
│   └── test/{emotion}/{sample_id}_ref.jpg
├── audio/
│   ├── train/{emotion}/{sample_id}.wav
│   ├── val/{emotion}/{sample_id}.wav
│   └── test/{emotion}/{sample_id}.wav
├── features/
│   └── audio_features/
│       ├── train/{emotion}/{sample_id}.npy
│       ├── val/{emotion}/{sample_id}.npy
│       └── test/{emotion}/{sample_id}.npy
├── text/
│   ├── annotations.csv
│   └── annotations.jsonl
├── metadata/
│   ├── dataset_manifest.csv
│   ├── dataset_manifest.json
│   ├── statistics.json
│   ├── reference_identity_stats.json
│   └── folder_validation_report.json
└── splits/
    ├── train_split.txt
    ├── val_split.txt
    └── test_split.txt
```

The `{emotion}` directory names are:

```text
angry, contemptuous, disgusted, fearful, happy, neutral, sad, surprised
```

## File Semantics

### `images/`

Target or evaluation images associated with the sample emotion label.

Example:

```text
images/test/angry/sample_xxx.jpg
```

This indicates that the target/evaluation image belongs to a test sample with the unified emotion label `angry`.

### `reference_images/`

Reference images provide the visual identity anchor for identity preservation.

Important: the emotion directory under `reference_images/` denotes the sample's target/control emotion, not necessarily the visible expression in the reference image. For example:

```text
reference_images/test/angry/sample_xxx_ref.jpg
```

This means the reference image is used as the identity anchor for a test sample whose target/control emotion is `angry`. The reference image itself may show a different expression.

### `audio/`

Emotion-aligned human speech waveforms. Audio provides an affective control signal and is aligned at the level of shared emotion semantics, not speaker identity.

### `features/audio_features/`

Precomputed audio features corresponding to the waveform files in `audio/`.

### `text/`

Text annotations in CSV and JSONL formats. These contain the emotion-conditioned text prompt and raw text fields.

### `metadata/`

Metadata and verification files.

`dataset_manifest.csv` and `dataset_manifest.json` are the authoritative sample indexes. Each row contains fields such as:

```text
sample_id
split
emotion
has_image
has_audio
has_text
has_reference_image
text
raw_text
valence
arousal
identity_id
alignment_scores
```

For the public release, original source paths, original source sample IDs, and source dataset identifiers are omitted.

### `splits/`

Plain-text files listing sample IDs for each split:

```text
train_split.txt
val_split.txt
test_split.txt
```

## Loading Protocol

Use `metadata/dataset_manifest.csv` as the authoritative index. Each row links the target image, reference image, audio waveform, precomputed audio feature, text annotation, emotion label, split, and reference-identity group.

The intended sample interpretation is:

```text
reference image -> identity condition
text/audio/emotion label -> affective control condition
target image -> supervision or evaluation target
```

## Leakage-Safe Evaluation

The benchmark enforces reference-identity-disjoint splits. Let each reference-identity group be defined by a shared visual reference image anchor. Then no group is allowed to appear in more than one split.

The final release satisfies:

- Cross-split reference-identity leakage: 0
- Near-duplicate overlap: 0
- Cross-dataset leakage: 0

This protocol ensures that validation and test evaluation measure generalization to held-out reference identities rather than memorization of identities seen during training.

## Benchmark Contribution

AffectHuman-43K is intended to support research on identity-preserving controllable generation under multimodal affective conditioning. Compared with standard affect-recognition datasets, image-editing datasets, or audio-only emotion datasets, AffectHuman-43K provides the following combination:

- complete image, reference-image, text, and audio coverage
- reference-image-based identity conditioning
- heterogeneous affective control through text, audio, and emotion labels
- repeated-reference groups for identity-consistency analysis
- reference-identity-disjoint train/validation/test splits
- verification metadata for leakage-safe evaluation

This makes the benchmark suitable for evaluating whether a model can preserve a visual identity while performing affective transformations specified through multiple modalities.

## License And Responsible Use

AffectHuman-43K is released under the [AffectHuman-43K Research-Only License](LICENSE). Access is limited to non-commercial academic research, benchmarking, and evaluation.

Users may not use the dataset for:

- commercial products, services, or deployments
- biometric identification or face-recognition deployment
- surveillance, tracking, profiling, or identity matching
- high-stakes emotion inference about real individuals
- impersonation, identity misuse, harassment, or deceptive synthetic media
- redistribution, mirroring, sale, sublicensing, or public rehosting of raw dataset files
- attempting to identify, re-identify, contact, or locate any individual represented in the dataset

Users may share aggregate research results, benchmark numbers, academic publications, and model evaluations, provided they do not expose raw dataset files, private identity information, or information that enables re-identification.

Users are responsible for complying with the licenses, terms, and usage restrictions of all upstream datasets, models, and source materials used with AffectHuman-43K. The AffectHuman-43K license does not grant rights beyond those permitted by applicable upstream licenses.

## Ethical Considerations

The AffectHuman-43K dataset was created to support research in identity-preserving affective controllable generation. Because the benchmark involves human faces, reference identity anchors, speech/audio, and emotion labels, users should treat the dataset as sensitive human-centered research data.

Recommended safeguards include:

- report evaluation results in aggregate
- avoid public release of identity-revealing examples unless explicitly permitted
- avoid use cases that could harm, target, profile, or deceive individuals
- document model limitations and biases when publishing results
- respect privacy and avoid any attempt to identify or contact individuals represented in the data

Access may be denied, revoked, or restricted if the maintainers determine that a requester has violated these terms or presents a misuse risk.

## Authors

- **Jamuna S. Murthy**  
  Ramaiah Institute of Technology

- **Amin Karimi Monsefi**  
  The Ohio State University

- **Rajiv Ramnath**  
  The Ohio State University

## Acknowledgments

We thank the contributors and institutions that supported the development of the AffectHuman-43K dataset.