New Kaggle competition: Image generation with diffusion models (www.kaggle.com)

Kaggle has launched its first text-to-image generation challenge focused on prompt-to-image alignment for diffusion models. Competitors must generate a single image per text prompt and are judged primarily on composition correctness using an F1 score that rewards faithful inclusion of expected objects. The contest includes 50 public prompts for local testing and a larger hidden set (150–200 prompts) for final scoring, pushing entrants to balance photorealism with strict prompt fidelity rather than optimizing for perceptual quality alone. The evaluation pipeline is fully automated: prompts are parsed with part-of-speech tagging to determine expected objects, and images are analyzed with object detection (YOLO) to identify present elements; the F1 score is computed from predicted vs. expected objects. Submissions require running and sharing a provided Kaggle Jupyter Notebook (for reproducibility) and uploading a generated submission.csv; organizers recommend DreamLayer — an open-source tool that exports images, results.csv, and optional config-dreamlayer.json in the required folder layout — though other workflows are allowed if they match the format. Rules prohibit renaming image files or altering results.csv. The competition offers modest cash prizes totaling $500 and serves as a practical benchmark to standardize alignment metrics and drive improvements in controllable, prompt-faithful image generation.