How High-Frequency Image Components Affect Generation Quality

A JPEG-Based Canonical Representation Approach

Master's Project at Albert-Ludwigs-Universität Freiburg
Computer Vision Group | Supervised by Arian Mousakhan

---

📋 Overview

This research investigates how high-frequency image components impact generation quality in two-stage generative models (VQGANs). By replacing learned tokenizers with deterministic JPEG compression, we demonstrate that aggressive frequency removal maintains generation performance despite reconstruction degradation.

Key Findings

• FID Scores: QF=20 (42.78) vs QF=40 (43.96) vs Baseline (43.92)
• Stable Generation: High-frequency loss doesn't degrade semantic understanding
• Training Stability: JPEG preprocessing eliminates codebook collapse

---

⚠️ Code Availability

The complete implementation is not publicly available due to:

• Proprietary research agreements
• Institutional privacy policies

This repository contains preliminary JPEG compression utilities only. But you can find architecture details in report attached below.

---

📦 Available Components

├── DCT_JPG.py              # JPEG DCT compression
├── compression.py          # Compression pipeline
├── huffman_parser.py       # Huffman coding utilities
├── quant.py                # Quantization analysis
└── output/                 # Sample outputs

---

Report - https://drive.google.com/file/d/11o6W0CFiWkaqf2uq4OGQDJObHr4CbU_n/view?usp=sharing The report includes:

• Complete methodology
• VQGAN+DINO architecture details
• Transformer training procedures
• Comprehensive experimental results
• Visual comparisons

---

🛠️ Requirements

pip install numpy opencv-python pillow

For full implementation (from report):

• PyTorch 2.0+
• timm (Vision Transformer)
• LPIPS, FID metrics
• BDD100K dataset

---

📊 Results Summary

Method	Embedding	FID ↓	rFID ↓
VQGAN	32-dim	40.66	20.53
VQGAN	16-dim	43.92	19.62
JPEG (QF=40)	16-dim	43.96	21.85
JPEG (QF=20)	16-dim	42.78	21.92

---

🎓 Citation

@mastersthesis{jadhav2025jpeg,
  title={How High-Frequency Image Components Affect Generation Quality: 
         A JPEG-Based Canonical Representation Approach},
  author={Jadhav, Sejal},
  year={2025},
  school={Albert-Ludwigs-Universit{\"a}t Freiburg},
  type={Master's Project}
}

---

📧 Contact

Sejal Jadhav
Supervisor: Arian Mousakhan
Examiner: Prof. Dr. Thomas Brox

---

🔗 Related Work

• VQGAN Paper - Esser et al., 2021
• DINO - Caron et al., 2021
• BDD100K Dataset

---

Computer Vision Group | University of Freiburg