Dataset Viewer
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
DREAM: Dialogue to REAlistic Multicultural Image Sequences
DREAM is a multicultural multimodal dataset linking persona-grounded dialogues with photorealistic portrait images and storyboard-like dialogue scene sequences.
The dataset was introduced in:
DREAM: A Multicultural Multimodal Dataset Linking Dialogues and Realistic Image Sequences
LREC 2026.
Dataset Overview
DREAM is a fully synthetic multimodal resource designed to support research on:
- visual grounding of dialogue
- identity consistency across image sequences
- appearance-based demographic perception
- multimodal evaluation and bias analysis
Each dialogue contains:
- Two enriched persona profiles
- structured demographic attributes
- narrative descriptions (appearance, personality, environment)
- Two profile portraits (one per speaker)
- Dialogue-level scene images aligned with visually salient turns
- Unified JSON schema linking text, personas, prompts, and images
All images are AI-generated and intended exclusively for research purposes.
Released Subset
This Hugging Face release contains approximately 75% of the full DREAM dataset:
- 750 dialogues
- 750 final dialogue JSON files
- 750 visual-turn specification JSON files
- 1500 profile portraits (PNG)
- 5190 dialogue scene images (PNG)
The subset includes:
- the full human-evaluated portion (250 dialogues)
- additional randomly selected dialogues
Dataset Structure
The released subset follows a structured layout combining JSON dialogue files, generated images, and evaluation metadata.
Each dialogue is represented by:
- a final JSON file containing persona profiles and dialogue text with image references
- a visual-turn specification file describing scene prompts and dialogue alignment
- profile portraits for both speakers
- dialogue scene images corresponding to visual turns
Dataset Contents
dream-75pct/
├── json/
│ ├── final/
│ │ ├── dialogue_id.json
│ │ └── ...
│ └── visual_turns/
│ ├── dialogue_id.json
│ └── ...
│
├── images/
│ ├── profiles/ # 1024x1024
│ │ ├── dialogue_id_A.png
│ │ ├── dialogue_id_B.png
│ │ └── ...
│ └── scenes/ # 1024x1024
│ ├── dialogue_id_img_1.png
│ ├── dialogue_id_img_2.png
│ └── ...
│
└── metadata/ # human evaluation results
└── evaluation.xlsx
JSON Format (Simplified)
Each dialogue file contains:
{
"dialogue_id": "...",
"profiles": {
"A": {
"profile_struct": {...},
"profile_narrative": {...},
"profile_prompt": "..."
},
"B": {...}
},
"dialogue": [
{"text_1": "..."},
{"text_2": "..."},
{"image_id": "..."}
]
}
Visual-turn files contain:
{
"dialogue_id": "...",
"visual_turns": [
{
"image_id": "...",
"dialogue_indices": [...],
"speaker_focus": "A|B|both",
"scene_type": "shared|memory|imagined|cutaway|montage",
"prompt": "..."
}
]
}
Example Dialogue Entry
Example excerpt from a dialogue JSON file:
{
"dialogue_id": "persona_chat_8310",
"profiles": {
"A": {
"profile_struct": {
"age": 59,
"gender_identity": "female",
"sexual_orientation": "heterosexual",
"ethnicity": "North African & Middle Eastern (MENA) (Egypt, Morocco, Saudi Arabia, Iran, etc.)",
"nationality": "United States of America",
"residence_country": "United States of America"
},
"profile_narrative": {
"personal_data": [
"My name is Laleh Farahani and I am a 59-year-old woman of Iranian descent who holds American citizenship.",
"I live on the west coast of the United States, where I have settled into a relaxed lifestyle near the ocean.",
"I work part-time in event catering, often picking up drinks for gatherings and social functions.",
"In my free time I spend hours indulging in comfort foods and binge-watching television series.",
"I am currently single and committed to improving my health through gradual lifestyle changes."
],
"visual_appearance": [
"I have thick, curly salt-and-pepper hair that I often pull back into a loose bun when I’m at home.",
"My skin has an olive tone reflecting my Middle Eastern heritage, and I have warm brown eyes that light up when I laugh.",
"I wear rectangular reading glasses with dark tortoiseshell frames, which I find both practical and stylish.",
"I favor comfortable, roomy clothing like soft sweaters and leggings, but I always add a quirky accessory like a colorful scarf.",
"I am on the shorter side with a curvy build that I’ve begun reshaping through portion control and light exercise."
],
"environment": [
"I live in a cozy one-bedroom apartment overlooking the Pacific Ocean, with a small balcony where I sip iced tea while watching the waves.",
"My living room is dominated by a large flat-screen TV and a plush couch that I practically live on during my binge-watching marathons.",
"The kitchen counters are always strewn with snack bowls and empty soda cans, evidence of my enthusiastic snacking habits.",
"I keep a mini bar cart nearby stocked with beers and mixers, ready for impromptu karaoke nights with friends.",
"A stack of workout DVDs and a set of resistance bands lurk in the corner as reminders of my ongoing effort to get in shape."
],
"personality_attitudes": [
"I’m a laid-back extrovert who loves a good laugh, even if it means pulling a harmless prank with a whoopee cushion now and then.",
"I’m self-deprecating but kind-hearted, often joking about my ‘gross’ habits to put others at ease.",
"I value camaraderie and genuine connections, and I’m always happy to pick up drinks or lift someone’s spirits.",
"Although I can be indulgent in my eating and TV habits, I’m committed to making steady progress toward healthier routines.",
"I believe life is too short to take too seriously, so I approach every evening out as an opportunity to make memories."
],
"other_details": [
"I have a collection of novelty whoopee cushions that I break out for quick pranks whenever the moment feels right.",
"My go-to karaoke song is an 80s power ballad that always gets the crowd singing along at the bar.",
"I have a habit of narrating cooking shows out loud, even when I’m microwaving frozen meals late at night.",
"Every Friday, I send out a group text for an informal bar meet-up and I’m famous for bringing everyone’s first round of drinks.",
"I once attempted to learn basic Spanish through a cooking class, but I reverted to English when the recipes got too complicated."
]
},
"profile_prompt": "59-year-old female, North African & Middle Eastern ethnicity, shoulders-up square portrait, neutral soft lighting, photorealistic with sharp facial focus. She has thick, curly salt-and-pepper hair loosely pulled into a bun, olive-toned skin and warm brown eyes that crinkle with a friendly smile. Rectangular dark tortoiseshell reading glasses perch on her nose, and she wears a cozy sweater layered with a colorful scarf draped casually around her shoulders. In the softly blurred background a hint of a sunny balcony railing and the Pacific Ocean’s waves can be seen."
},
"B": {
"profile_struct": {
"age": 35,
"gender_identity": "female",
"sexual_orientation": "gay/lesbian",
"ethnicity": "North American (USA, Canada, Greenland, etc.)",
"nationality": "United States of America",
"residence_country": "United States of America"
},
"profile_narrative": {
"personal_data": [
"My name is Emma Carter, and I’m a 35-year-old female teacher and sports coach in the United States of America.",
"I hold American citizenship and live in a suburban community not far from the city limits.",
"I have a bachelor’s degree in elementary education and additional certification in girls’ sports coaching.",
"I am single and deeply invested in my work with sixth graders and in coaching the girls’ volleyball team after school.",
"Outside of school hours, I host karaoke nights at a local bar every Friday to blend my love of music with community engagement."
],
"visual_appearance": [
"I have straight, shoulder-length blonde hair that I wear down or pull into a simple ponytail during practice sessions.",
"My skin is fair with a light sprinkling of freckles across my cheeks and the bridge of my nose.",
"I often wear sporty-rimmed glasses when grading papers, then switch to sleeker frames for my karaoke performances.",
"My typical attire includes athletic leggings or track pants with a team polo when coaching, and casual jeans or a flowy blouse for evenings out.",
"I have an athletic build and average height for a woman, with toned arms from years of demonstrating drills."
],
"environment": [
"I live in a modest brick house with a small backyard where I hang laundry and host weekend barbecues for friends.",
"My classroom is decorated with student artwork, motivational posters, and neatly labeled bins of supplies.",
"At the community center gym, I always have my whistle, a set of stopwatches, and a stash of volleyballs in my coaching bag.",
"On bar nights, I set up on a small stage under colored lights, complete with a microphone stand and a ready playlist of karaoke tracks.",
"My kitchen counter often holds a bowl of fruit—apples and pears—despite my mild allergy to certain varieties."
],
"personality_attitudes": [
"I’m energetic and nurturing, finding deep satisfaction in helping young students learn and grow.",
"I have a playful side that comes out during karaoke, where I belt out songs with enthusiastic abandon.",
"I’m patient by nature, though I’m honest when my sixth graders push boundaries with pranks or volleyball antics.",
"I value community involvement, which is why I combine teaching, coaching, and social evenings into a cohesive lifestyle.",
"I maintain a pragmatic approach to health, enjoying fresh fruit despite occasional allergic reactions and balancing treats with nutritious meals."
],
"other_details": [
"I carry a detailed coaching planner everywhere, filled with practice drills, game strategies, and motivational quotes.",
"I challenge my students to crossword puzzles during downtime to keep their critical thinking sharp.",
"I crochet small matching team scarves as end-of-season gifts for my volleyball players.",
"My signature rallying cry is “Alright, team!” which I use to motivate both students and karaoke participants alike.",
"I once organized a school talent show fundraiser where I performed a duet with the principal on stage."
]
},
"profile_prompt": "35-year-old female, North American ethnicity, shoulders-up square portrait, neutral soft lighting, photorealistic with sharp facial focus. She has straight shoulder-length blonde hair tied in a low ponytail, fair skin dotted with a light sprinkling of freckles, and sporty-rimmed glasses framing her attentive eyes. She wears a team polo shirt with a whistle on a lanyard around her neck, suggesting her role as a coach. The softly blurred background offers a subtle suggestion of a gym environment with an out-of-focus volleyball net and stopwatches."
}
},
"dialogue": [
{
"persona_id": "persona_chat_8310_A",
"text_1": "Are you going out tonight?"
},
{
"persona_id": "persona_chat_8310_B",
"text_2": "Hi. What is with kids these days? I am fun but they do not agree."
},
{
"image_id": "persona_chat_8310_img_1"
},
{
"persona_id": "persona_chat_8310_A",
"text_3": "It always a nice time old or not."
},
{
"persona_id": "persona_chat_8310_B",
"text_4": "My sixth graders are really getting out of hand lately."
},
{
"persona_id": "persona_chat_8310_A",
"text_5": "Well I cant party with kids but nice job it seems."
},
{
"persona_id": "persona_chat_8310_B",
"text_6": "I coach too but the girls throw their volley balls at me some times. On purpose!"
},
{
"image_id": "persona_chat_8310_img_2"
},
{
"persona_id": "persona_chat_8310_A",
"text_7": "Are whoopee cushions still fun?"
},
{
"image_id": "persona_chat_8310_img_3"
},
{
"persona_id": "persona_chat_8310_B",
"text_8": "No. It reminds me of all the pranks they pull on me. Do you like kids?"
},
{
"persona_id": "persona_chat_8310_A",
"text_9": "Yes but throw the ball at them?"
},
{
"persona_id": "persona_chat_8310_B",
"text_10": "I am too nice. What other teacher holds karaoke friday parties?"
},
{
"image_id": "persona_chat_8310_img_4"
},
{
"persona_id": "persona_chat_8310_A",
"text_11": "I am too fat to move I am slimming down now."
},
{
"image_id": "persona_chat_8310_img_5"
},
{
"persona_id": "persona_chat_8310_B",
"text_12": "Good for you. Try some apples, I am allergic but eat the ones they give me."
},
{
"image_id": "persona_chat_8310_img_6"
},
{
"persona_id": "persona_chat_8310_A",
"text_13": "Yea fine I will come sing really loud."
},
{
"persona_id": "persona_chat_8310_B",
"text_14": "Thank goodness. We do crosswords and crochet at the same time."
},
{
"persona_id": "persona_chat_8310_A",
"text_15": "I will bring beer for when the kids leave."
},
{
"persona_id": "persona_chat_8310_B",
"text_16": "Sounds like a plan. We can watch the football game when they are gone too."
},
{
"image_id": "persona_chat_8310_img_7"
}
]
}
Data Generation Pipeline
DREAM was produced through a controlled synthetic pipeline:
- Dialogue sourcing (PersonaChat + ComperDial)
- Persona profile expansion
- Portrait rendering
- Visual-turn extraction
- Scene rendering
- Human + automated evaluation
Models used:
- text LLM: OpenAI o4-mini
- text-to-image: OpenAI gpt-image-1
- automated evaluation: vision-language model OpenAI GPT-5
Evaluation
Human evaluation includes:
- realism
- dialogue-image coherence
- character consistency
- appearance-based demographic perception
A subset of dialogues was triple-rated, enabling inter-annotator agreement analysis.
Intended Use
DREAM is intended for research on:
- multimodal dialogue grounding
- visual consistency benchmarking
- synthetic dataset analysis
- demographic perception studies
- multimodal generation evaluation
Limitations
- All images and personas are synthetic.
- Appearance-based demographic labels are operational categories for research only.
- The dataset should NOT be used for real identity inference or surveillance.
Ethical Considerations
DREAM is a fully synthetic dataset designed for research on multimodal dialogue and identity representation.
All personas and images are generated and do not correspond to real individuals.
Demographic attributes refer to visual appearance categories rather than real-world identity claims and should not be interpreted as ground truth demographic information.
The dataset should not be used for surveillance, biometric identification, or any system making real-world identity judgments.
Acknowledgments
DREAM builds upon dialogue data from the following publicly available datasets:
- PersonaChat
- ComperDial
We thank the creators of these resources for making their datasets publicly available, which enabled the construction of DREAM.
License
This dataset is released under:
CC-BY-NC-4.0
You may use, modify, and redistribute the dataset for non-commercial research purposes with attribution.
Citation
If you use DREAM in your research, please cite:
@inproceedings{mallo2026dream,
title={DREAM: A Multicultural Multimodal Dataset Linking Dialogues and Realistic Image Sequences},
author={Mallo de la Calle, Juan and Estecha-Garitagoitia, Marcos and Córdoba, Ricardo and D'Haro, Luis Fernando},
booktitle={Proceedings of the LREC 2026 Conference},
year={2026}
}
Related Resources
GitHub repository (pipeline and scripts):
https://github.com/JuanMallodelaCalle/personalized-ai-visual-summaries
Contact
Juan Mallo de la Calle
📧 juan.mallo@alumnos.upm.es
Speech Technology and Machine Learning Group
Universidad Politécnica de Madrid
- Downloads last month
- 226