Diffusion Models: Pioneering Efficiency, Control, and Unlearning in Generative AI
Latest 100 papers on diffusion models: Oct. 17, 2026
Diffusion models have rapidly become a cornerstone of generative AI, pushing the boundaries of what’s possible in image, video, language, and scientific data synthesis. Recent research highlights a concerted effort to enhance their efficiency, imbue them with fine-grained control, ensure safety, and even understand their inner workings—from memorization to feature dynamics. This digest unpacks some of the latest breakthroughs, showcasing how innovation is driving these models towards unprecedented capabilities and practical applications.
The Big Idea(s) & Core Innovations
A central theme emerging from recent papers is the pursuit of efficiency without compromise. Traditional diffusion models can be computationally intensive, but novel approaches are tackling this head-on. For instance, DaTMe (Dynamics-aware Token Merging) by SeungJu Cha et al. from Hanyang University, South Korea introduces a training-free framework that significantly accelerates image generation by intelligently merging tokens based on diffusion dynamics, rather than text relevance alone. This preserves structural saliency and image-wide details, addressing a key limitation of prior merging methods that overlooked crucial visual information. Similarly, ManiFusion by Minkyu Kim et al. from Seoul National University proposes a “superposition in manifold space” to process multiple samples within a single denoising evaluation, achieving nearly 2x throughput improvement with minimal quality degradation. This relies on the insight that the low-dimensional structure of clean data manifolds allows multiple samples to coexist and retain distinct denoising paths in a shared representation.
Beyond raw speed, advancements in fine-grained control and adaptive generation are expanding the utility of diffusion models. The EOS-Story framework by SeungJu Cha et al. from Hanyang University, Seoul, South Korea revolutionizes visual storytelling by leveraging the <EOS> embedding in CLIP to carry phrase-level appearance information, ensuring subject identity consistency while adapting to diverse scene contexts. This enables rich, multi-attribute appearance cues to be infused without semantic dilution. For image stylization, Amir Semmo’s work, Local Content-Style Control for Diffusion-based Image Stylization, introduces
Share this content:
Discover more from SciPapermill
Subscribe to get the latest posts sent to your email.
Post Comment