Visual Compositional Generation Research Application
This is a shared research work between Dr. Soleymani (MLL) and Dr. Rohban (RIML Lab) from Sharif University of Technology on
Visual Compositional Generation in Diffusion-based Text-to-image Models with a goal of submitting a
survey on the related problems in a 3-month milestone.
Project Description:
Large-scale diffusion-based models have achieved state-of-the-art results on text-to-image (T2I) synthesis tasks. Despite their significant ability of generating high quality and prompt-aligned images, these models are exposed to a number of major compositional-related drawbacks such as
object missing and
improper attribute binding.
Hence, we aim to study and categorize the compositional-related failure modes and the corresponding methods to overcome the problems.
For more information you can read the following papers.
- T2I-CompBench: A Comprehensive Benchmark for Open-world Compositional Text-to-image Generation
- Attend-and-Excite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models
- Training-Free Structured Diffusion Guidance for Compositional Text-to-Image Synthesis
The Musts:
- Being familiar with Linear Algebra fundamentals is necessary.
- Being familiar with Statistics and Probability fundamentals is necessary.
- Being familiar with Machine Learning and Deep Learning fundamentals is necessary.
- Being familiar with Transformers (Attention Mechanism) and Diffusion Models is
recommended (Note that there will be some short training sessions on the mentioned topics during the project).
- Dedicating considerable time to the project is necessary.
Process:
- Deadline of submitting the application: 10/24/2023 - 23:59 (Tehran Time)
- First wave of rejection/acceptance: 10/25/2023 - 23:59 (Tehran Time)
- Short Interviews: 10/26/2023 - 23:59 (Tehran Time)
- Final wave of rejection/acceptance: 10/27/2023 - 23:59 (Tehran Time)
Please fill this form till 10/24/2023 - 23:59 (Tehran Time). Note that the deadlines are somehow harsh and it's important to dedicate as much time as possible.
https://docs.google.com/forms/d/e/1FAIpQLSfRSteyD6y0LUF2RI4GiMa3Bv_N8w_EI5OTSt3F0j1teAMHSA/viewform?usp=sf_link
We would be happy to answer any questions you may have through
[email protected].
@sutmll
#research_application