sportsvision/Describe-then-Generate-Bottleneck

The Describe-then-Generate-Bottleneck dataset investigates the phenomenon of information loss when AI systems generate images from textual descriptions of existing images. This dataset contains 150 randomly selected samples that demonstrate the "bottleneck" effect occurring in describe-then-generate pipelines.

The dataset explores a critical question in AI vision: How much visual information is lost when we ask a vision-language model to describe an image, and then use that description to generate a new image?

Dataset Structure

This dataset contains paired images and text demonstrating the describe-then-generate bottleneck:

1. Original Prompt → Generated Image A (direct generation)
2. Generated Image A → VLM Description → Generated Image B (bottleneck generation)

Data Fields

• prompt_id: Unique identifier for each sample
• original_prompt: The original text prompt used for image generation
• prompt_generated_image: Image generated directly from the original prompt using Gemini-2.5-Flash-Image-Preview
• vlm_description: Detailed description of the prompt-generated image created by Gemini-2.5-Flash VLM
• description_generated_image: Image generated from the VLM description using Gemini-2.5-Flash-Image-Preview

Data Instances

Each instance represents a complete describe-then-generate cycle:

{
  "prompt_id": "sample_001",
  "original_prompt": "A serene mountain landscape at sunset with a lake reflecting orange and pink clouds",
  "prompt_generated_image": <PIL.Image>,
  "vlm_description": "The image shows a tranquil mountain scene during golden hour...",
  "description_generated_image": <PIL.Image>
}

Data Collection Process

1. Source Selection: 150 samples were randomly selected from the zai-org/ImageRewardDB dataset
2. Image Generation: Original prompts were used to generate images using Gemini-2.5-Flash-Image-Preview
3. Vision-Language Description: Generated images were processed through Gemini-2.5-Flash VLM to create detailed descriptions
4. Bottleneck Generation: VLM descriptions were used to generate new images, completing the bottleneck pipeline
5. Quality Control: All generations were manually verified for completeness and quality

Models Used

• Image Generation: Google Gemini-2.5-Flash-Image-Preview
• Vision-Language Model: Google Gemini-2.5-Flash
• Source Dataset: zai-org/ImageRewardDB

Use Cases

This dataset is valuable for research in:

• Information Loss Analysis: Quantifying how much visual information is lost in describe-then-generate pipelines
• Vision-Language Model Evaluation: Assessing the quality of image descriptions and their faithfulness to original content
• Multimodal AI Benchmarking: Creating benchmarks for vision-language-vision model chains
• Semantic Preservation Studies: Understanding which visual elements are preserved vs. lost in VLM descriptions
• AI Art and Generation Research: Studying the bottleneck effects in creative AI applications

Citation

@dataset{describe_generate_bottleneck_2025,
  title={Describe-then-Generate-Bottleneck: A Dataset for Studying Information Loss in Vision-Language-Vision Pipelines},
  author={SportVision Research Team},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/sportsvision/Describe-then-Generate-Bottleneck}
}