Paper Short Showcase

We present a quantization-aware image enhancement pipeline that preserves quality after deployment by training the model to handle quantization noise directly. The method combines gated encoding, multi-scale refinement, and fake quantization during training to support mobile-friendly inference with strong visual fidelity.

We fine-tune a vision-language model for traffic video captioning and question answering by separating spatial and temporal reasoning, then filtering frames to better capture the most relevant context in each scenario.