SyntaGen

Computer vision has been rapidly transformed by advancements in generative models, particularly in text-to-image generation with models like Imagen 3, Stable Diffusion 3, Flux, and DALLE-3, as well as text-to-video models such as Sora, Stable Video Diffusion, and Meta MovieGen. In the realm of 3D generation, models like Zero-123, Instant 3D, and the Large Reconstruction Model (LRM) have pushed the boundaries of 3D content creation. These innovations have enabled the development of highly realistic and diverse synthetic visual datasets, complete with annotations and rich variations, which are invaluable for training and evaluating algorithms in object detection, segmentation, representation learning, and scene understanding. The second SyntaGen Workshop aims to foster collaboration and knowledge exchange across the field, bringing together experts and practitioners to propel the development of generative models and synthetic visual datasets to new heights. Through talks, paper presentations, poster sessions, and panel discussions, the workshop will catalyze breakthroughs at the intersection of generative models and computer vision applications.

Speakers

TBD

Schedule

TBD

Accepted Papers

Call for Papers

Topics

The main objective of the SyntaGen workshop is to offer a space for researchers, practitioners, and enthusiasts to investigate, converse, and cooperate on the development, use, and potential uses of synthetic visual datasets made from generative models. The workshop will cover various topics, including but not restricted to:

Leveraging pre-trained generative models to generate data and annotations for perception-driven tasks, including image classification, representation learning, object detection, semantic and instance segmentation, relationship detection, action recognition, object tracking, and 3D shape reconstruction and recognition.
Extending the generative capacity of large-scale pre-trained text-to-image models to other domains, such as videos, 3D, and 4D spaces.
Exploring new research directions in generative models, including GANs, VAEs, diffusion models, and autoregressive models, to advance visual content generation.
Synergizing expansive synthetic datasets with minimally annotated real datasets to enhance model performance across scenarios including unsupervised, semi-supervised, weakly-supervised, and zero-shot/few-shot learning.
Enhancing data quality and improving synthesis methodologies in the context of pre-trained text-to-image (T2I), text-to-video (T2V), text-to-3D, and text-to-4D models.
Evaluating the quality and effectiveness of the generated datasets, particularly on metrics, challenges, and open problems related to benchmarking synthetic visual datasets.
Ethical implications of using synthetic annotated data, strategies for mitigating biases, verifying and protecting generated visual contents, and ensuring responsible data generation and annotation practices.

Important workshop dates

Organizers


Khoi Nguyen VinAI Research, Vietnam	Anh Tuan Tran VinAI Research, Vietnam	Binh Son Hua Trinity College Dublin, Ireland	Supasorn Suwajanakorn VISTEC, Thailand	Yi Zhou Adobe

SyntaGen - Harnessing Generative Models for Synthetic Visual Datasets