Diffusion models are powerful, but evaluating them consistently is difficult. Some images capture prompts accurately but look unrealistic, while others look great but ignore key details.

This challenge asks you to balance both:
- Make images faithful to the text prompts.
- Ensure they contain the expected objects.

##Data

- [Public prompts](https://docs.google.com/document/d/1VJHTKDMPapyLbACAIOWvzRdrUhV70y03Mbyr_y--AN0/edit?usp=sharing): 50 prompts for local testing and debugging.
- Private prompts: 150–200 prompts used for scoring (hidden).

##Submission Format

To be considered on the leaderboard, participants must complete two steps:

1 - Run and Share the Jupyter Notebook
- You must run the provided [JupyterNotebook](https://github.com/DreamLayer-AI/DreamLayer/blob/main/KaggleDreamLayer.ipynb) directly on the Kaggle platform for reproducibility. 
- Update the first cell with your own drive link which contains a folder with the images and the report generated via [DreamLayer](https://github.com/DreamLayer-AI/DreamLayer) 
- Share the notebook with the competition hosts. Sharing is mandatory. Submissions without a shared notebook will not be counted.
- The notebook will automatically display your F1 score at the end of its run.


2 - Submit the "submission.csv" file 
- Along with sharing the notebook, upload the "submission.csv" file via Kaggle’s submission system.
- This file is automatically generated at the end of the notebook’s run. It contains the predicted outputs required for scoring.

Drive Folder structure:

drive-link-folder/
├─ 0001.png
├─ 0002.png
├─ ...
├─ results.csv   # maps prompt_id to predictions
└─ config-dreamlayer.json   # optional run metadata

**Important rules:**
- Do not rename the generated image files.
- Do not modify results.csv (this is generated automatically by DreamLayer).
- The Kaggle Notebook must be shared with organizers for your submission to count.

##How to participate

Submissions can be generated locally using DreamLayer, an open-source benchmarking tool. DreamLayer gives you full control over model choice, parameters, and reproducibility, and it automatically exports results in the required format (`images/`, `results-dreamlayer.csv`, optional `config-dreamlayer.json`).

[DreamLayer GitHub](https://github.com/DreamLayer-AI/DreamLayer)

We recommend DreamLayer for easy reproducibility and streamlined exports, but participants are free to use any workflow or tool of their choice as long as submissions follow the required format.

Running the  [JupyterNotebook](https://github.com/DreamLayer-AI/DreamLayer/blob/main/KaggleDreamLayer.ipynb) on Kaggle Platform is compulsory along with submitting the "submission.csv" file for reproducibility.