image-to-image translation

Image-to-image translation transforms images from one visual domain to another while preserving structure. **Examples**: Sketch to photo, day to night, summer to winter, horse to zebra, photo to painting, map to satellite. **Approaches**: **Paired training**: pix2pix requires aligned source/target pairs, learns direct mapping. **Unpaired training**: CycleGAN learns from unpaired examples using cycle consistency loss. **Modern diffusion**: SDEdit, img2img add noise then denoise toward target domain. **Key architectures**: Conditional GANs, encoder-decoder networks, cycle-consistent adversarial training. **Diffusion img2img**: Start from encoded input image + noise, denoise with text conditioning toward new domain. Denoising strength controls how much original is preserved. **Applications**: Photo editing, artistic stylization, domain adaptation, synthetic data, virtual try-on, face aging. **Style-specific models**: GFPGAN (face restoration), CodeFormer, specialized checkpoints. **Challenges**: Preserving identity/structure across transformation, handling diverse inputs, artifacts. Foundational technique enabling countless creative and practical applications.