Visual Insight Assistant

Turn a photo into a caption, then compare it with human descriptions from Flickr8k.

Captioning model: ViT-GPT2. Review generated descriptions for mistakes and assumptions. Only upload photos you have permission to use. Your uploaded photo is processed in your browser by this app.

Choose a Flickr8k photo or upload your own image.

Flickr8k test photo Or upload a photo Five human reference captions

Generated description

The first run downloads the model and can take several minutes. Keep this tab open. Later runs use the cached model.

Model caption Draft alt text for human review Rule based keywords

Evaluation

Generate captions for the first ten Flickr8k test images. Each score compares one caption with five human references using smoothed sentence BLEU4.

Photo	Generated caption	BLEU4	Seconds

Evaluation results will appear here.

BLEU measures word overlap, not accuracy. A correct description may use different words from the references. Caption time includes image preparation and excludes model loading. This small sample does not establish general performance, and possible overlap with training data remains unknown.

Data: jxie/flickr8k. Model: Xenova/vit-gpt2-image-captioning. The app runs with Transformers.js in your browser on a static Hugging Face Space. Check the original dataset terms before redistributing photos.