Core Concepts
Non-isotropic diffusion model and innovative entropy model improve image compression quality.
Abstract
Abstract:
Non-isotropic diffusion model enhances image quality by distinguishing frequency contents.
Novel entropy model accurately models latent representation probability distribution.
Introduction:
Learning-based methods surpass classical codecs in rate-distortion performance.
Generative-based codecs aim for realistic reconstructions.
Related Works:
Diffusion models offer stable training and high-quality image generation.
Methods:
Blurring diffusion model improves image quality through distinct schedules.
Proposed entropy model efficiently encodes latent representation into a binary stream.
Experiments:
Merged dataset used for training with various hyperparameters tested.
Comparison with SOTA Methods:
Our method shows superior performance in rate-perception tradeoff but lags in distortion compared to other methods.
Visual Quality:
Our model achieves high-quality reconstructions with fewer artifacts compared to other models.
Ablation study:
Maximum blurring level impacts reconstruction quality significantly.
Laplacian-shaped positional encoding results in notable bitrate savings compared to other encoding types.
Stats
モデルは2.4百万ステップで最適化されました。
初期学習率は1 × 10^-4から1 × 10^-7まで段階的に減少しました。
λの値は{0.0004,0.005,0.01,0.02,0.04,0.016}から選択されました。
Quotes
"Non-isotropic diffusion model enhances perceptual quality by distinguishing between frequency contents."
"Our proposed framework yields better perceptual quality compared to cutting-edge generative-based codecs."