Linearly transformed color guide for low-bitrate diffusion based image compression.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Diffusion models generate high-quality images but lack precise global color control without fine-tuning.
• Controlling color aspects in generative models is crucial for applications requiring specific color palettes.

Purpose of the Study:

• To develop a method for controlling the global color of images generated by diffusion models without requiring model retraining or fine-tuning.
• To improve the fidelity and realism of compressed images using novel color guidance techniques.

Main Methods:

• Rewriting diffusion model guidance equations to steer generation towards a target color map.
• Introducing a compression framework that combines semantic and general color information for low-cost decoding.
• Analyzing the scaling of guidance during the diffusion process.

Main Results:

• The proposed method achieves precise global color control in generated images without compromising generation quality.
• The color guidance scaling should increase throughout the diffusion process for optimal results.
• The method significantly improves the fidelity and realism of images compressed at extremely low bit rates (10⁻² bpp).

Conclusions:

• The novel guidance equations offer effective control over image color in diffusion models.
• The integrated compression framework demonstrates superior performance in low-bitrate image compression compared to classical and semantic methods.
• This research provides a practical solution for color-controlled image generation and efficient image compression.