Efficient, Training-Free Single-Image Diffusion Model

This proposes a new method that generates images while preserving the internal structure (patch distributions across various scales) of a single reference image. Existing Single-image diffusion approaches required training a separate model for each single image, which demanded massive computational cost and hours of optimization time.

This research solves this problem by modeling images as a patch dataset across various scales. Leveraging the fact that patch dimensions are small and the dataset is finite, it computes the score function of noisy patches using an optimal closed-form denoiser. This enables an efficient image diffusion model to be implemented without any neural network training process.

Key features and achievements are as follows:

• Training-free: Instant image generation is possible without any separate optimization process.
• High performance: Achieved state-of-the-art (SOTA) generation quality and diversity compared to existing training-based Single-image diffusion models.
• Diverse applications: Applicable to unconditional image generation, text-guided stylization, image symmetrization, retargeting, and more.
• Overwhelming speed: Compatible with Latent space diffusion, and provides acceleration technology capable of generating megapixel-scale images within 1 second and gigapixel-scale images within a few minutes.