Efficient and Training-Free Single-Image Diffusion Models
6 hours ago
- #diffusion models
- #single-image generation
- #patch-based denoising
- The paper introduces a training-free method for generating images by matching the internal patch distribution of a single reference image.
- Instead of training a diffusion model, it uses a finite dataset of patches at multiple scales and applies an optimal closed-form denoiser to compute the score function.
- The approach achieves state-of-the-art quality and diversity, surpassing trained single-image diffusion models.
- Applications include unconditional image generation, text-guided stylization, image symmetrization, and retargeting.
- It is compatible with latent space diffusion and includes acceleration techniques for fast megapixel and gigapixel generation.