Mohamed Elawady. 2025. Advancing Assessment Practices in CS Education through AI-Generated
Visual Test Cases. In UK and Ireland Computing Education Research Conference (UKICER 2025),
September 04–05, 2025, Edinburgh, United Kingdom. ACM, New York, NY, USA, 1 page.
https://doi.org/10.1145/3754508.3754538
-
Image Classification Validator:
- Used models: ResNet18, ConvNextV1-Tiny
- Pre-trained weights: ImageNet
- Function: Upload an AI-generated image to see top-3 predictions from classification models with probabilities.
- [Google Colab]
- [HF Spaces Demo]
-
Image Detection Validator:
- Used models: YOLOv8n, YOLOv12n
- Pre-trained weights: COCO
- Function: Upload an AI-generated image to view the top three predictions from detection models, including their probabilities, along with the corresponding bounding boxes for each model.
- [Google Colab]
- [HF Spaces Demo]
- generate realistic image (cat in room) and write bounding box coordinates of cat within the image to be used later to compare with object detection models. don't show annotation box over the generated image.
- Data Folder (images and detection information in json format file): ./imgs/chatgpt-gpt5/
- Link: https://chatgpt.com/
- Discussion: Successfully generated images with bounding box information and text descriptions in JSON format. Further example generation is constrained by paywall restrictions.
- Data Folder (images and text of LLM reply): ./imgs/copilot-gpt5/
- Link: https://m365.cloud.microsoft/chat/
- Discussion: Capable of generating images and text with highlighted detection bounding boxes, but lacks support for exporting results as downloadable files and performs with slow execution speed.
- Data Folder (images): ./imgs/gemini-2.5-flash/
- Link: https://gemini.google.com/
- Discussion: Capable of generating images; however, automatic bounding box generation for object detection is not supported.
- Data Folder (images): ./imgs/stable-diffusion-2.1/
- Links:
- Discussion: Cannot generate annotation metadata for object detection as it is restricted to text-to-image generation. Outputs tend to be less realistic (mainly focusing on the same body pose) and rely on simple prompts such as 'cat in room,' though performance is superior to Stable Diffusion 2.1.
- Data Folder (images): ./imgs/stable-diffusion-2.1/
- Links:
- Discussion: Unable to generate annotation metadata for object detection since the model is limited to text-to-image generation. The outputs are less realistic, and only basic prompt text such as 'cat in room' can be used.
- Data Folder (images): ./imgs/dall.e-mini/
- Link: https://huggingface.co/spaces/dalle-mini/dalle-mini
- Discussion: One of the earliest open-source text-to-image generation models, lacking object detection metadata. The outputs are highly unrealistic.
- Data Folder (screenshot of LLM reply): ./imgs/claude-sonnet4/
- Link: https://claude.ai/chat/
- Discussion: Unable to automatically generate images or annotation metadata for object detection. However, capable of creating and running an annotation application that allows manual image upload, bounding box drawing, and exporting annotations in multiple formats (COCO, YOLO, Pascal VOC).