Opening and Welcome

Photographic Style Editing with Diffusion Prior
— Towards Artificial Intelligence Generated Photography (AIGP)

• [Abstract]

Photographic style editing modifies photographic concepts such as focus, aperture, tone, and lighting while preserving the original content of an input image. This talk will present two advancements in photographic style editing by leveraging the diffusion prior. To render the lens blur effect of fast apertures, we propose BokehDiff, which employs a physics-inspired self-attention module to align with the image formation process with a one-step diffusion model and overcomes the rendering artifacts in depth discontinuities with the diffusion prior. To transfer the personal photographic style of professional photographers, we introduce PIF (Personalized Image Filter), which learns the average photographic concepts (e.g., vignetting, contrast, shadow, etc.) with the diffusion prior and achieves the personalized photographic style with a content-preserving residual denoising paradigm. We hope the realistic photographic style editing results from our works could take one step closer towards Artificial Intelligence Generated Photography (AIGP).

Boxin Shi

Peking University

Zero-Shot *Blind* Super-Resolution using “Internal Learning”

• [Abstract]

WBlind-SR methods aim to be agnostic to the unknown downscaling SR-kernel with which the low-resolution (LR) image was generated. However, even advanced Blind-SR methods are still fundamentally restricted to rather simplistic downscaling SR-kernels (e.g., anisotropic Gaussian kernels), and fail on more complex (out of distribution) downscaling kernels. In this talk I will show how Zero-Shot Blind-SR can overcome this limitation, by exploiting the strong recurrence of tiny patches inside a single natural image (the LR test image alone). In the first part of my talk I will review our series of works on Zero-Shot Blind-SR using such “Internal Learning”, and the rationale behind it. I will then present “KernelFusion” – a new Zero-Shot Diffusion-based Blind-SR method which simultaneously recovers the unique image-specific SR-kernel and  its corresponding high-resolution (HR) image. This is obtained by:  (i) training an image-specific patch-based diffusion model on the single LR input image alone, thus capturing its unique internal patch statistics; and (ii) exploiting the principle that the correct SR-kernel is the one that maximizes patch similarity between the HR and LR images. By breaking free from predefined distributions of training datasets, KernelFusion pushes Blind-SR into a new assumption-free paradigm, handling complex downscaling kernels previously thought impossible.

Michal Irani

Weizmann Institute of Science

Effective Image and Video Enhancement for Mobile Devices

• [Abstract]

In this talk, I will review our work on image/video enhancement, frame interpolation, and related topics. The underlying theme is to learn from image data for these tasks based on classic approaches (e.g., depth, optical flow, and image pyramids) and recent learning methods (e.g., transformers). Specifically, I will present our recent work on segmentation, grounding, inpainting, and restoration, all of which can be applied to mobile devices. When time allows, I will also discuss our work for other vision tasks.

Ming-Hsuan Yang

University of California, Merced & Google DeepMind

Efficient Visual Generation on the Edge

• [Abstract]

Generative visual models achieve stunning results but remain too compute-intensive for widespread deployment. This talk introduces recent work that tackles this challenge head-on, including 1) efficient tokenizer and token compression: deep compression auto-encoders (DC-AE) for diffusion models, which compress tokens up to 64× to cut cost,  2) use linear attention: SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers and  3) exploit sparsity: Radial Attention: O(n log n) Sparse Attention with Energy Decay for Long Video Generation and Sparse VideoGen: Accelerating Video Diffusion Transformers with Spatial-temporal Sparsity 4) exploit quantization and nvfp4:  SVDQuant: Absorbing Outliers by Low-Rank Components for 4-Bit Diffusion Models. Together, these approaches—token compression, linear attention, sparse attention, quantization—point the way toward making high-fidelity image and video generation not only cheaper and faster, but also practical on edge devices.

Song Han

Massachusetts Institute of Technology

Closing