theme-worlds

Theme Worlds is a re-implementation of MultiDiffusion for Stable Diffusion v1.5 that generates seamless 360° panoramas. The panoramas have left and right edges that are identical, creating a continuous loop. The method uses stock SD v1.5 weights without any training or source images, as the loop is a property of the sampler rather than a post-processing step. The project includes an interactive viewer where users can drag to look around and scroll or pinch to zoom. It reports a mean seam ratio of 1.12 over 28 panoramas, where 1.0 means the join is like any other pair of columns. It also notes that circular windows produce a 7.6× smaller seam than without, based on 12 paired runs. The approach composes themes rather than concatenating them, as CLIP silently drops 20 tokens from a naive two-theme prompt. The project was built by Marlen Melis for KAIST AIC.20000 Introduction to AI Computing (Fall 2026). The images are the notebook's own PNG outputs, re-encoded as JPEG, and the page wraps them around a cylinder.