Datasets:

vatsal-malaviya
/

AcT2I-Prompts

license: cc-by-4.0
dataset_info:
  features:
    - name: id
      dtype: string
    - name: base_prompt
      dtype: string
    - name: animal1
      dtype: string
    - name: animal2
      dtype: string
    - name: action
      dtype: string
    - name: rarity_label
      dtype: string
    - name: emotional_valence
      dtype: string
    - name: spatial_topology
      dtype: string
    - name: temporal_extent
      dtype: string
    - name: emotional_prompt
      dtype: string
    - name: spatial_prompt
      dtype: string
    - name: temporal_prompt
      dtype: string
  splits:
    - name: train
      num_bytes: 94793
      num_examples: 125
  download_size: 51438
  dataset_size: 94793
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - text-to-image
language:
  - en
tags:
  - T2I
  - Reasoning
  - Action
  - Benchmark
size_categories:
  - n<1K

AcT2I-Prompts

What is this?

AcT2I-Prompts is the core prompt set from the AcT2I benchmark. It contains 125 base action-centric prompts that describe interactions between two animal agents (e.g. "a goose competing for dominance with a turkey") plus 3 enriched variants per base prompt:

spatial_prompt
emotional_prompt
temporal_prompt

Each base prompt is also labeled along semantic axes:

rarity_label (how biologically/common the interaction is)
emotional_valence (aggressive / defensive / affiliative / communicative)
spatial_topology (pursuit vs physical contact vs distant interaction)
temporal_extent (instantaneous vs extended action)

Total rows: 125 (one per base prompt). Each row includes all 4 textual variants.

This repo intentionally does not include generated images or human study results. Those are released separately.

Intended use

This dataset is for evaluation / analysis of text-to-image models, not for training.

Typical use:

For each row, take the base_prompt and (optionally) the enriched spatial_prompt, emotional_prompt, and temporal_prompt.
Generate images from your T2I model for each variant.
Measure whether the model's image actually depicts the described interaction and action.

These prompts are meant to stress-test spatial, temporal, and affective reasoning ("who is doing what to whom, in what posture, with what intent, at what moment").

Out-of-scope / disallowed use

This dataset is not intended for:

Training or promoting violent / graphic animal content for shock or harassment.
Generating deceptive media presented as "real" wildlife attacks or staged cruelty.
Drawing conclusions about human social behavior, human interpersonal violence, or human identity bias. The benchmark is deliberately animal–animal and two-agent focused.

Do not use this dataset to build abusive content pipelines.

Data fields

Each row in data/prompts.jsonl represents one base interaction scenario.

id (int)
base_prompt (str)
animal1 (str)
animal2 (str)
action (str)
rarity_label (str: frequent | rare | very_rare)
emotional_valence (str: aggressive | defensive | affiliative | communicative)
spatial_topology (str: proximal-contact | pursuit / avoidance | distant interaction)
temporal_extent (str: instantaneous | extended action)
spatial_prompt (str)
emotional_prompt (str)
temporal_prompt (str)

There are no train/dev/test splits. All 125 rows are considered the official evaluation set.

Dataset creation

Curation rationale

Most existing "compositional" prompts test simple attribute binding ("a blue cat on a skateboard"). AcT2I instead targets interaction semantics: chasing, comforting, retaliating, asserting dominance, surrendering, etc. These require:

asymmetric roles (one agent acts on the other),
physically plausible contact / pursuit / restraint poses,
temporal cues (in the middle of an attack vs after being struck),
emotional / intent cues (aggressive vs affiliative).

We focus on animal–animal interactions (instead of human–human violence or human identity scenarios) to:

Reduce sensitive social/ethical risk around representing harm between humans.
Get clearer signal about action depiction instead of immediately running into "the model can't draw human hands" failures.

How prompts were generated

We defined pairs of animals and an interaction verb (e.g. "competing for dominance with", "comforting", "chasing", "retaliating against").
We wrote a concise base_prompt for each interaction.
For each base prompt, we produced three enriched variants:
- spatial_prompt: adds explicit body orientation / physical layout.
- emotional_prompt: adds affect / intent wording.
- temporal_prompt: anchors the scene in a specific moment or phase of action.
We assigned semantic labels (rarity_label, emotional_valence, spatial_topology, temporal_extent) to each base prompt.

Who created the data

All prompts, enriched variants, and semantic labels were authored/verified by the AcT2I team. No personal names, locations, or other PII were included.

Bias, risks, and limitations

Violence / aggression content: Many prompts explicitly describe aggression, dominance, pursuit, or threat between animals. This is intentional (models struggle most with these high-contact, asymmetric actions). However, it means the dataset can be used to generate violent-looking content. Please use responsibly.
Scope limitations: The benchmark is animal–animal only and two-agent only. Results should not be overgeneralized to human social interactions, medical scenarios, multi-agent scenes, tool use, etc.
Biological plausibility: Some interactions are biologically rare or borderline impossible. That is deliberate: we care about whether the model can depict the requested interaction clearly, not whether the interaction is common in nature.

Citation

If you use AcT2I-Prompts, please cite:

@article{malaviya2025act2i,
  title={AcT2I: Evaluating and Improving Action Depiction in Text-to-Image Models},
  author={Malaviya, Vatsal and Chatterjee, Agneet and Patel, Maitreya and Yang, Yezhou and Baral, Chitta},
  journal={arXiv preprint arXiv:2509.16141},
  year={2025}
}