FreeMD: Training-free multi-domain text-to-image generation with any control.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Diffusion models advance controllable text-to-image generation.
• Existing methods struggle with semantic alignment and structure consistency due to coupled controls and limited domain focus.

Purpose of the Study:

• To introduce FreeMD, a training-free method for improved text-to-image generation.
• To enhance semantic alignment with text prompts and achieve superior structure consistency.

Main Methods:

• Proposed FreeMD, a novel training-free multi-domain text-to-image generation method.
• Introduced independent appearance and structure guidance branches.
• Employed multi-domain guidance combining spatial and frequency domains for structure control.

Main Results:

• FreeMD achieves better semantic alignment with text prompts.
• The method demonstrates excellent structure consistency with control signals.
• Experiments show FreeMD outperforms existing methods in controllability and generation quality.

Conclusions:

• FreeMD effectively decouples text and structure controls for improved text-to-image generation.
• The multi-domain guidance strategy enhances both semantic accuracy and structural coherence.
• FreeMD offers a plug-and-play solution for various generative models and downstream tasks.