The Second Workshop on Efficient and On-Device Generation (EDGE) at CVPR 2025 will focus on the latest advancements of generative AI in the computer vision domain, with an emphasis on efficiencies across multiple aspects. We encourage techniques that enable generative models to be trained more efficiently and run on resource-constrained devices, such as mobile phones and edge devices. Through these efforts, we envision a future where these permeating generative AI capabilities become significantly more accessible with virtuous scalability and plateauing carbon footprint.

The topics involved in the workshop include but are not limited to:

• Modeling efficiency:
• Design of efficient modules and algorithms for generative models.
• Efficient architecture for generative models: auto-encoders, diffusion, GAN, autoregressive models, etc.
• Efficient transformers and state-space models.
• Training efficiency:
• Methods for reducing the memory footprint of training generative models.
• Using knowledge distillation to support model compression.
• Hardware-aware training acceleration and energy efficient computing.
• Online training under edge constraints.
• Parameter efficient tuning.
• Mixed pre-training to enable downstream tasks via lightweight tuning.
• Inference efficiency:
• Techniques for reducing the computational complexity of generative models, such as quantization, sparsification, and MoEs.
• Advanced sampling techniques for fast inference.
• Approaches for efficient deployment and distribution generative models across multiple devices or edge devices.
• Hardware-aware inference acceleration and energy efficient computing.
• Data efficiency:
• Few-shot diffusion model.
• Generative model finetuning.
• Training generative models with fewer data.
• Application-specific efficiency:
• Applications of efficient generative AI in computer vision, such as image / video generation, image / video editing, style transfer, and super-resolution.
• Efficient personalization and post-generation editing.
• Efficient generative models for real-time video synthesis: talking head, vitual try-on, gaming engine, etc.

Format: Submissions must use the CVPR 2025 Author Kit for Latex/Word Zip file and follow the CVPR 2025 author instructions and submission policies. Submissions need to be anonymized. The workshop considers two submission tracks:

• Long Paper: submissions are limited to 8 pages excluding references;
• Extended Abstract: submissions are limited to 4 pages excluding references.

Only long papers will be included in the CVPR 2025 proceedings.

Submission Site: TBD

Submission Deadline: March 21, 2025 (AOE)

Workshop Date: TBD

Workshop Location: TBD