RIML Lab

💠 Compositional Learning Journal Club

Join us this week for an in-depth discussion on Compositional Learning in the context of cutting-edge text-to-image generative models. We will explore recent breakthroughs and challenges, focusing on how these models handle compositional tasks and where improvements can be made.

✅ This Week's Presentation:

🔹 Title: Divide, Evaluate, and Refine: Evaluating and Improving Text-to-Image Alignment with Iterative VQA Feedback

🔸 Presenter: Amir Kasaei

🌀 Abstract:
Recent advancements in text-conditioned image generation, particularly through latent diffusion models, have achieved significant progress. However, as text complexity increases, these models often struggle to accurately capture the semantics of prompts, and existing tools like CLIP frequently fail to detect these misalignments.

This presentation introduces a Decompositional-Alignment-Score, which breaks down complex prompts into individual assertions and evaluates their alignment with generated images using a visual question answering (VQA) model. These scores are then combined to produce a final alignment score. Experimental results show this method aligns better with human judgments compared to traditional CLIP and BLIP scores. Moreover, it enables an iterative process that improves text-to-image alignment by 8.7% over previous methods.

This approach not only enhances evaluation but also provides actionable feedback for generating more accurate images from complex textual inputs.

📄 Paper: Divide, Evaluate, and Refine: Evaluating and Improving Text-to-Image Alignment with Iterative VQA Feedback


Session Details:
- 📅 Date: Sunday
- 🕒 Time: 2:00 - 3:00 PM
- 🌐 Location: Online at vc.sharif.edu/ch/rohban


We look forward to your participation! ✌️

www.tg-me.com/jp/telegram/com.RIMLLab/133

2.0K viewsArash Marioriyad, edited  Sep 14, 2024 at 07:58

💠 Compositional Learning Journal Club

Join us this week for an in-depth discussion on Compositional Learning in the context of cutting-edge text-to-image generative models. We will explore recent breakthroughs and challenges, focusing on how these models handle compositional tasks and where improvements can be made.

✅ This Week's Presentation:

🔹 Title: Divide, Evaluate, and Refine: Evaluating and Improving Text-to-Image Alignment with Iterative VQA Feedback

🔸 Presenter: Amir Kasaei

🌀 Abstract:
Recent advancements in text-conditioned image generation, particularly through latent diffusion models, have achieved significant progress. However, as text complexity increases, these models often struggle to accurately capture the semantics of prompts, and existing tools like CLIP frequently fail to detect these misalignments.

This presentation introduces a Decompositional-Alignment-Score, which breaks down complex prompts into individual assertions and evaluates their alignment with generated images using a visual question answering (VQA) model. These scores are then combined to produce a final alignment score. Experimental results show this method aligns better with human judgments compared to traditional CLIP and BLIP scores. Moreover, it enables an iterative process that improves text-to-image alignment by 8.7% over previous methods.

This approach not only enhances evaluation but also provides actionable feedback for generating more accurate images from complex textual inputs.

📄 Paper: Divide, Evaluate, and Refine: Evaluating and Improving Text-to-Image Alignment with Iterative VQA Feedback


Session Details:
- 📅 Date: Sunday
- 🕒 Time: 2:00 - 3:00 PM
- 🌐 Location: Online at vc.sharif.edu/ch/rohban


We look forward to your participation! ✌️

BY RIML Lab