Consistent Style Generation in AIGC: Style Aligned and Story Diffusion

Recent advances in AIGC image generation have shown great value in e‑commerce and content creation. Beyond basic prompt‑based generation, techniques such as ControlNet and IP‑Adapter enable spatial control and style transfer.

This article introduces two representative methods for producing multiple images with consistent style: the Style Aligned method and the Story Diffusion method. Style Aligned achieves style consistency by sharing a self‑attention mechanism across a batch of images, allowing each image to attend to the features of the first image in the batch.

Story Diffusion operates in two stages. First, a training‑free Consistent Self‑Attention module generates a series of images with coherent semantics. Second, a transformer block predicts intermediate frames in a latent space, which are then decoded into video frames.

Experimental results on home‑decoration scenarios demonstrate that both methods can maintain overall style across different viewpoints, though fine‑grained spatial structure and detail alignment still need improvement.

The discussion concludes with a summary of the methods, their current limitations, and future research directions for coherent multi‑image and video generation.