Shared attention keeps SDXL from drifting between styles

What it does The repository contains notebooks and demos for a Google research technique that generates batches of images sharing a single visual style using SDXL, or copies a style from a reference photo. It also combines with ControlNet-Depth and MultiDiffusion to add depth conditioning or generate style-consistent panoramas with SD V2.

The interesting bit The clever part is treating attention maps as a shared resource across the batch to enforce a common aesthetic. The code also treats this as a modular layer: it slots into existing pipelines alongside ControlNet and MultiDiffusion without replacing the underlying model.

Key highlights

• Implements style alignment via shared attention for SDXL and Stable Diffusion V2.
• Supports style transfer from a reference image.
• Composes with ControlNet-Depth and MultiDiffusion for conditioned or panoramic output.
• Includes local browser demos and an online Hugging Face Space for ControlNet + StyleAligned.
• Explicitly flagged as not an officially supported Google product.

Caveats

• Two items on the roadmap are still open: multi-style support with MultiDiffusion and integration with DreamBooth.
• The code was validated only on a specific environment: Python 3.11, PyTorch 2.1, and Diffusers 0.16.

Verdict Worth exploring if you are tinkering with diffusion internals or need a quick way to enforce style consistency across a batch. Give it a pass if you need a finished, officially supported tool.