Encoder–Decoder Portfolio · Project 02

Vision Encoder-Decoder Image Captioning

Upload an image, compare greedy and beam-search captions, adjust caption length, inspect a caption-structure heuristic and genuine exported token trace, and review model limitations.

Runtime: not loaded Static browser app No paid API

Responsible use

Captions may be incomplete, biased, generic, or incorrect. The browser structure score is not a correctness probability. Human review is required before accessibility-critical use. Do not upload private, confidential, medical, identity, workplace, or copyrighted images without permission.

1 · Image

Choose an image

Safe synthetic samples

2 · Controls

Caption settings

The first model load downloads ONNX files into the browser cache. Later runs are usually faster.

Greedy caption

—

Choose an image and generate captions.

Caption structure heuristic—

Beam-search captions

—
  1. Beam candidates will appear here.

Model-generated alt-text draft

Human review required

A short, low-repetition beam candidate will appear here. Human review is required.

Model comparison

Not run

Configure a fine-tuned Transformers.js model ID in metadata.json to enable live comparison.

3 · Interpretability

Token-generation view

Bundled sample images can show genuine token order and transition scores exported from the Python model. Uploaded images are clearly marked unavailable because the browser pipeline does not expose token probabilities.

StepTokenScorePartial caption
Generate a caption to populate this view.

4 · Evidence

Evaluation and model details

Evaluation status

Architecture

Known limitations