PaperlayerView exact revision on arXiv

Source-linked AI summary

HumanCentric-Scenes-AI: A Human-Curated Multimodal Dataset of AI-Generated Human-Centric Scenes Across Surveillance, Occupational Portraits, AI-Edited Images, and Short-Form Video

Nima Kamali, Sana Nasiri, Negar Kamali

arXiv:2610.29045v1cs.CVcs.AIcs.HC

TL;DR

Research on AI-generated media has often focused on face manipulation or whole-image generation, while human-centric contexts include compressed surveillance, occupational portraits, partial edits, and short-form video. HumanCentric-Scenes-AI releases manually curated stimuli and per-item metadata across those four domains. Its analyses report domain-specific regularities, including sharper-than-real CCTV-style images and gender skew under neutral occupational prompts.

  • Problem

    AI-generated depictions of people span compressed surveillance, occupational portraits, partial image edits, and short-form video, but detection research has focused on other media settings.

    Source: §1 · pp. 1–2

  • Method

    The authors release 296 manually curated stimuli across four domains, with prompts where recorded and per-stimulus metadata.

    Source: §1 · p. 2, §3.1 · p. 3

  • Results

    Across analyses, CCTV-style images remain sharper than real frames, 73.9% of gender-neutral occupational generations depict men, larger edits associate with higher zero-shot accuracy, and videos span six subject categories.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Released metadata lets researchers test the reported regularities on the full subsets.

    Source: §8 · p. 8

    Several statistics use 25-stimulus characterization subsets, and difficulty estimates may age as generators improve.

    Source: §8 · p. 8

Abstract

from arXiv · show

We introduce HumanCentric-Scenes-AI, a human-curated dataset of 296 AI-generated stimuli that depict people across four contexts in which people encounter AI-generated media. The Surveillance subset contains 85 AI-generated CCTV-style images. The Occupations subset contains 128 AI-generated occupational portraits in a balanced paired design covering 64 occupations with one female and one male version of each. The AI-edited subset contains 27 real photographs in which a region was altered or replaced by a generative model. The Video subset contains 56 AI-generated short videos. We curated the images and videos manually over multiple rounds to make them as photorealistic as possible. For each stimulus we release the file, the generation prompt where recorded, and per-stimulus metadata. We report depicted-gender distributions from an initial set of 134 generations without a specified gender and compare them with U.S. Bureau of Labor Statistics occupational gender shares. We report edit-region size and its association with zero-shot detection difficulty for the AI-edited subset. We report image-quality distributions for the Surveillance subset and subject-category coverage for the Video subset. HumanCentric-Scenes-AI is intended first for human-perception research, and also for calibration of AI-detection classifiers and bias analysis of generative models.

1 Introduction

HumanCentric-Scenes-AI provides human-curated stimuli across four contexts where people encounter AI-generated depictions of people. The paper reports domain-specific findings on image realism, gender representation, edit size, and video coverage.

1 Introduction

  • The four contexts reflect distinct settings: compressed surveillance, occupational portraits with gender associations, partially edited photographs, and short-form social video.The cited motivation notes that detector traces weaken under compression and resizing, while generative models can amplify gender associations in occupational portraits [1].Source: §1 · pp. 1–2
  • The release contains 296 stimuli: CCTV-style images, paired portraits for 64 occupations, AI-edited photographs, and short videos, curated manually for photorealism with prompts where recorded and per-stimulus metadata.The paired portrait design includes one female and one male version of each occupation.Source: §1 · p. 2
  • Midjourney CCTV-style images remain sharper than real frames; 73.9% of gender-neutral occupational generations depict men, descriptive phrasing softens the skew, and larger edits associate with higher zero-shot accuracy.The characterized videos span six subject categories, with median durations similar to their real references.Source: §1 · p. 2
  • An initial set of 134 generations without specified gender reveals a generator default that the balanced release set cannot show.The analysis compares depicted-gender distributions with U.S. Bureau of Labor Statistics occupational gender shares.Source: §1 · p. 2

2 Related Work

Prior detection datasets emphasize face manipulation or whole-image generation, while other work addresses partial edits and demographic bias. HumanCentric-Scenes-AI is positioned as a small, metadata-rich, human-perception stimulus set spanning four domains.

2 Related Work

  • Face-swap and reenactment benchmarks manipulate faces while leaving the rest of the scene fixed, whereas whole-image datasets target detector generalization across generators.GenImage pairs over a million generated images with real counterparts across eight generators; Community Forensics covers several thousand models.Source: §2 · p. 2
  • For partial manipulation, TGIF provides roughly 80,000 text-guided inpainting forgeries, while this paper's small curated AI-edited set is intended for human-perception research.In the AI-edited subset, larger edits are associated with higher zero-shot classification accuracy.Source: §2 · p. 3
  • Prior bias studies quantify demographic representation and occupational stereotypes; this dataset adds a complementary paired occupational design that separates prompt phrasing from the generator's male default.The paper reports 82.1% male depictions under occupation-name prompts and a 16.4-percentage-point prompt-phrasing effect.Source: §2 · p. 3
  • The dataset differs in purpose from large detection benchmarks: it is a small, structured-metadata release curated as stimulus material for human-perception research.It follows the datasheet reporting framework of Gebru et al..Source: §2 · p. 3

3 Dataset Composition

The dataset is organized into four named stimulus subsets. Curators manually selected candidates for photorealism, and the release provides a metadata schema with one row per file.

3 Dataset Composition

  • The release is organized as four named subsets, with Table 1 giving headline counts for the full release.The supplied passage does not state the table's individual counts.Source: §3 · p. 3

3.1 Curation procedure

  • Curators reviewed candidate generations over multiple rounds, selecting images without immediately distinguishing artifacts; subset-specific prompts and paired designs shaped the stimuli.The Surveillance and Occupations images use Midjourney; videos use Veo, Sora, and Kling and were developed against real reference clips in matching subject categories.Source: §3.1 · pp. 3–4

3.2 Metadata schema

  • The metadata schema is presented in Table 2, with one row for each released file.The table lists the columns of metadata.csv.Source: Table 2 · p. 4, §3.2 · p. 4

4 Low-Resolution and CCTV Images

The Surveillance subset contains 85 Midjourney CCTV-style images, while image-quality characterization covers a 25-image subset and paired real CCTV frames. Despite prompts aimed at CCTV degradation, the generated images were sharper on the reported measure.

4 Low-Resolution and CCTV Images

  • The 85 images were prompted to mimic CCTV noise, blur, and downsampling; the characterization compares 25 stimuli with 26 real frames.Prompted Midjourney versions were v7 for 23 images, v6.1 for 6, and v6 for 3; 53 prompts specified no version.Source: §4 · p. 4

4.1 Composition and metadata

  • Image-quality statistics cover the 25-image experimental subset and paired real references, not the full 85-image release.The released analysis script can be extended to the full release with a one-line change.Source: §4.1 · p. 4, §4 · p. 4

4.2 Generation prompts

  • A representative prompt requests grainy, low-resolution CCTV footage with compression, digital noise, blur, and low frame rate.The example specifies a parking security camera and Midjourney v6.Source: §4.2 · p. 4

4 Low-Resolution and CCTV Images

  • Median Laplacian variance was 649 for Midjourney images versus 212 for real CCTV frames in the 25-stimulus comparison.The table reports Mann–Whitney U = 482, z = 3.29, p = 0.001.Source: Table 3 · p. 5

5 Images with Gender–Occupation Pairings

The Occupations release pairs female- and male-subject portraits across 64 occupations, while a separate 134-generation set measures Midjourney’s gender default under gender-neutral prompts. That initial set shows a male skew, with descriptive role phrasing softening it, and female depictions under occupation-name prompts generally aligning with female-majority U.S. occupations.

5 Images with Gender–Occupation Pairings

  • The released Occupations subset pairs one female and one male portrait for each of 64 occupations, so its gender balance is by construction.The unconstrained gender prior is instead observable in the separate initial generation set.Source: §5 · p. 5

5.1 Gender distribution under gender-neutral prompts

  • Midjourney depicted a man in 73.9% of 134 generations without specified gender, including 82.1% under occupation-name prompts and 65.7% under role descriptions.The initial set covered 67 occupations and is not part of the release.Source: §5.1 · p. 5

5.2 Gender-stereotype congruence

  • Of 11 occupations rendered female under occupation-name prompts, seven of the eight with published BLS shares are female-majority occupations.The table reports that seven of eight published shares exceed the 47.1% overall employed-women baseline; Newscaster is 41.1% [18].Source: §5.1 · p. 5, Table 5 · p. 6
  • The study defines stereotype congruence by whether depicted gender matches the occupation’s female- or male-majority direction in U.S. labor statistics.Among the 11 occupations rendered female, seven were female-majority, Newscaster was male-majority, and three had no published BLS share.Source: §5.2 · p. 6

5.3 Prompt-template ablation

  • Fifteen of 67 occupations changed depicted gender between prompt styles; 13 flipped toward female under descriptive phrasing and two toward male.The two-sided sign test yielded p = 0.007.Source: §5.3 · p. 6

6 Partially AI-Generated Images

The AI-edited subset comprises 27 partially manipulated photographs, with paired analyses on 25 aligned source-image pairs. Edit-region size is positively associated with zero-shot classification accuracy, while the altered-pixel share varies by edit type.

6 Partially AI-Generated Images

  • The subset contains 27 Midjourney inpainting derivatives of real photographs, with pairwise analyses limited to 25 pairs having aligned source photographs.Inpainting can remove, modify, or add content while leaving the rest of the scene intact.Source: §6 · p. 6
  • Person-to-person replacement was the most common edit (9 of 25), followed by background replacement (6) and object insertion or replacement (5).Four edits replaced a person with an object or animal.Source: §6 · pp. 6–7
  • Background and person-to-object-or-animal edits altered the largest pixel shares, while person-to-person replacements altered a median 12.7%.The surrounding scene remains in person-to-person replacements.Source: §6 · p. 7

6.1 Edit-region size vs. detection difficulty

  • Edit-region size was positively associated with zero-shot classification accuracy across 25 images (Spearman ρ = 0.55, permutation p = 0.005).Three language models were evaluated; edits altering a median 14.3% of pixels spanned outcomes from 0 to 3 models correct.Source: §6.1 · p. 7

7 Everyday Social Videos

The Video subset contains 56 AI-generated short-form videos across six subject categories, produced with Veo, Sora, and Kling. Technical characterization covers 20 stimuli, and a representative Sora prompt describes a handheld, naturally lit home-kitchen pasta tutorial.

7 Everyday Social Videos

  • The Video subset contains 56 short-form videos across six subject categories, generated with Veo, Sora, and Kling.Per-video measurements cover a 20-stimulus characterization subset.Source: §7 · p. 7, Table 9 · p. 8

7.1 Sora prompt complexity

  • The representative Sora prompt combines a point-of-view pasta lesson in a home kitchen with handheld camera and natural lighting.The prompt is framed as a YouTube-video style scene featuring a man teaching.Source: §7.1 · p. 8

8 Discussion

The discussion summarizes patterns across the four domains and notes that some characterizations use only 25 stimuli, while difficulty estimates can change as generators improve.

8 Discussion

  • The dataset analyses report sharper-than-real CCTV-style images, gender associations under neutral prompts, and higher zero-shot accuracy for larger edits.The Video subset spans six subject categories, and characterized videos have median durations similar to real references.Source: §8 · p. 8
  • Some statistics use 25-stimulus characterization subsets, and difficulty estimates age as generators improve.The same scripts can be applied to full subsets, and recorded generator identities support future measurement of that decay.Source: §8 · p. 8
  • The AI-edited subset uses edited regions over public-source photographs, while the Surveillance, Occupations, and Video subsets depict synthetic people.The released materials are under CC BY-NC 4.0, and the dataset must not be used to train generative models.Source: §8 · pp. 8–9
  • AI-generated videos and real references both have median durations of 8.0 s, while AI-generated videos span a wider range of aspect ratios.Source: Table 10 · p. 9

A Datasheet

The datasheet presents the dataset as human-perception research material and describes its composition, collection, uses, licensing, and maintenance; it follows Gebru et al.'s [6] datasheet framework.

A Datasheet

  • The dataset contains 296 stimuli across four named subsets, with per-stimulus metadata and synthetic people in three subsets.The AI-edited subset instead derives from public-source photographs with an altered region and no intentionally added personally identifying information.Source: A Datasheet · p. 11
  • Surveillance and Occupations images were generated with Midjourney, while videos used Veo, Sora, and Kling and AI-edited stimuli modify regions of public-source photographs.Occupations uses paired female and male versions for 64 occupations; files are renamed to content hashes, with original descriptive names retained in metadata.Source: A Datasheet · p. 11
  • The dataset is intended for human-perception research and also supports classifier calibration, generative-model bias analysis, and prompt-template ablation.Training generative models and uses implying depicted synthetic people are real are prohibited.Source: A Datasheet · p. 11
  • The AI-generated assets, metadata, and curation are hosted on Hugging Face under CC BY-NC 4.0.Source: A Datasheet · p. 11
  • Midjourney versions are recorded for only 32 of 85 Surveillance images, and Veo stimuli are pending repository restoration with generator labels needing correction.Source: A Datasheet · p. 11

B Example Stimuli

The examples show how occupational prompt wording relates to depicted gender, how edit size relates to zero-shot accuracy, and what the four stimulus domains look like.

B Example Stimuli

  • Across 67 occupations, Midjourney defaults to male depictions, while descriptive prompts soften that default and occupation-name prompts render women more often in female-majority occupations.The figure contrasts prompts naming the occupation with prompts describing the role without naming it.Source: Figure 1 · p. 12
  • Edit-region size is positively associated with zero-shot classification accuracy (Spearman ρ = 0.55, permutation p = 0.005).Each point represents one of 25 AI-edited images; accuracy is averaged across GPT-4.1, Claude Opus, and Gemini Pro.Source: Figure 2 · p. 13
  • The examples show Surveillance, Occupations, AI-edited, and Video stimuli in that order, and all shown stimuli are AI-generated.Source: Figure 3 · p. 14
Loading 2610.29045v1…