Training-Free Single-Image Diffusion Models
18 hours ago
- #single-image modeling
- #patch-based generation
- #training-free diffusion
- Proposes a training-free diffusion model for generating images that match the internal patch structure of a single reference image.
- Uses a closed-form denoiser on patches across scales, eliminating neural network training and enabling efficient, scalable image generation.
- Accelerates generation via FlashAttention, latent space downsampling, and patch clustering to achieve megapixel speeds and gigapixel scalability.
- Demonstrates applications in unconditional generation, text-guided stylization, image retargeting, symmetrization, and tileable image creation.
- Shows compatibility with latent diffusion and maintains style-structure alignment by constraining output to input patches.