Taming Outlier Tokens in Diffusion Transformers

Apple Research Blog
Read full post
Researchers identified and addressed the issue of outlier tokens in Diffusion Transformers used for image generation, which degrade performance by corrupting local patch semantics. They proposed Dual-Stage Registers (DSR) to mitigate these outliers, improving generation quality across ImageNet and text-to-image tasks.

More in Image Generation

OpenAI launches ChatGPT Images 2.5, which it says cuts image generation latency by up to 50% vs. Images 2.0, and adds a Sketch feature for drawing in ChatGPT (Ina Fried/Axios)

Covered by 4 sources
Image Generation9 min read

Rebuilding AUTOMATIC1111 with Gradio Workflow

Hugging Face