Prompt2SceneBench: Structured Prompts for Text-to-Image Generation in Indoor Environments

Prompt2SceneBench is a structured prompt dataset with 12,606 text descriptions designed for evaluating text-to-image models in realistic indoor environments. Each prompt describes the spatial arrangement of 1–4 common household objects on compatible surfaces and in contextually appropriate scenes, sampled using strict object–surface–scene compatibility mappings.

A usecase of the Prompt2SceneBench has been showcased in the Prompt2SceneGallery image dataset which has been generated using SDXL and the prompts from Prompt2SceneBench dataset.

Dataset Structure

CSV Format (prompt2scene_prompts_final.csv)

Size: 12,606 prompts

Each row in the CSV corresponds to a single prompt instance and includes the following fields:

type: Prompt category — one of A, B, C, or D, based on number of objects and complexity.

object1, object2, object3, object4: Objects involved in the scene (some may be None/NaN/Null depending on type).

surface: The surface where the objects are placed (e.g., desk surface, bench).

scene: The indoor environment (e.g., living room, study room).

prompt: The final structured natural language prompt.

Note:

Type A prompt has only 1 object (object2, object3, object4 fields will be None/NaN/Null)

Type B prompt has only 2 objects (object3, object4 fields will be None/NaN/Null)

Type C prompt has only 3 objects (object4 field will be None/NaN/Null)

Type D prompt has 4 objects (all the object fields will have values)

Sample Examples:

Type A: a football located on a bench in a basement. (object1: football, surface: bench, scene: basement)

Type B: a coffee mug beside a notebook on a wooden table in a home office. (object1: coffee mug, object2: notebook, surface: wooden table, scene: home office)

Type C: a jar, a coffee mug, and a bowl placed on a kitchen island in a kitchen. (object1: jar, object2: coffee mug, object3: bowl, surface: kitchen island, scene: kitchen)

Type D: An arrangement of an air purifier, a pair of slippers, a guitar, and a pair of shoes on a floor in a bedroom. (object1:air purifier, object2: pair of slippers, object3: guitar, object4: pair of shoes, surface: floor, scene: bedroom)

JSON Format (prompt2scene_metadata.json)

The JSON contains the following keys:

objects: List of all 50 objects used in the prompt generation.

scenes: List of 15 indoor scenes.

surfaces: List of 20 compatible surfaces.

object_to_scenes: Dictionary mapping each object to plausible indoor scenes.

object_to_surfaces: Dictionary mapping each object to compatible surface(s).

surface_to_scenes: Dictionary mapping each surface to scene(s) where it naturally occurs.

prompt_templates: Template used for generating the prompts for all the prompt types (A, B, C and D), each prompt type has 3 variants

This JSON file supports reproducibility and reuse by providing all internal mappings used during structured prompt generation. The community can further extend/modify the above lists and mappings and use their own prompt templates based on the usecase.

Prompt2SceneBench can be directly used for:

Prompt-to-image generation using models like Stable Diffusion XL to benchmark compositional accuracy in indoor scenes.

Prompt–image alignment scoring, evaluating how well generated images match the structured prompts.

Compositional generalization benchmarking, testing models on spatial arrangement of 1–4 objects with increasing difficulty.

Zero-shot captioning evaluation, using prompts as pseudo-references to measure how captioning models describe generated images.

Scene layout reasoning tasks, e.g., predicting spatial configuration or scene graph generation from textual prompts.

Style transfer or image editing tasks,where the structured prompt can guide object placement or scene modification in indoor contexts.

Multimodal fine-tuning or distillation, where paired structured prompts and generated images can be used to improve alignment in vision-language models (VLMs), especially for grounding objects, spatial relationships, and indoor scene context.

Controllable generation studies, analyzing prompt structure impact on generated outputs under different text-to-image models.