Details

Type

  • Bachelor Thesis
  • Master Thesis

Persons

1

Quantifying Texture Compression Quality with Vision-Language Models

Image quality metrics such as PSNR, LPIPS and FLIP measure the error a compression algorithm introduces by condensing it into a single number. Each has known weaknesses. PSNR measures pixel-wise differences and ignores how humans actually perceive images. Perceptual metrics such as LPIPS and FLIP are tuned to approximate human judgment, but they can reward or penalize features a human observer would not notice. The most reliable alternative, a user study, is expensive and slow, and rarely practical for every comparison in a paper.

Recent work suggests that large language models can stand in for human participants in survey-style studies. Maier et al. (arXiv:2510.08338) show that LLMs reproduce human ratings closely when they answer in free text that is then mapped onto a rating scale, rather than being asked for a number directly.

This thesis investigates whether vision-language models (VLMs), such as Claude, ChatGPT or open-weight alternatives, can deliver reliable and reproducible quality scores for compression artifacts when prompted and set up correctly. Key questions include:

  • How should the model be queried (absolute ratings vs. pairwise comparisons against a reference) to get consistent results?
  • How well do VLM scores agree with human judgments, compared to PSNR, LPIPS and FLIP?
  • How reproducible are the scores across runs, prompts and model versions?

Tasks

  • Compress a set of textures using common methods 
  • Implement a framework that shows a VLM a reference image and its compressed versions, and turns the model's responses into quality scores.
  • Check whether the scores behave sensibly.

Requirements

  • Knowledge of English language (source code comments and final report should be in English)
  • Knowledge of texture compression algorithms is advantageous, but not necessary
  • More knowledge is always advantageous.

Responsible

For more information please contact Elias Kristmann.