SyncDiffusion
SyncDiffusion is an official open-source implementation of the NeurIPS 2023 research project titled Coherent Montage via Synchronized Joint Diffusions. Developed by researchers from the KAIST Geometric AI Lab, this software serves as a plug-and-play module designed to solve the problem of visible seams and scene incoherence often found when stitching images into panoramas using standard diffusion models. The system works by synchronizing multiple diffusion processes through gradient descent derived from a perceptual similarity loss. At each denoising step, it calculates gradients based on predicted denoised images to provide meaningful guidance, ensuring that overlapping regions blend seamlessly while maintaining scene consistency. The package supports various generation modes including loop-closed generation and conditional generation using ControlNet. It includes evaluation tools, a command-line interface for panorama creation, and a Gradio-based web demo for interactive use with custom prompts. The softwar