SILMM-Implementation

Last Update: 2025.01.26; All codes are implemented by my team including me.

This is UNOFFICIAL code implementation for the paper:

"SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation (2024.12)"

📍 Notice

1. I only implemented for SEED-LLaMA model (Discrete MLLM).
2. By training only 1 iteration, the paper results are reproduced. (Authors trained for 3 iterations.)

📍 5-Step (code implementation)

• 1. Compositional Prompt Generation
• 2. Diverse Image Generation
• 3. Decompositional Self-Questioning
• 4. VQA-based Self Feedback
• 5. Learning from Self-Feedback

📍 Baseline Model

Please refer to this original SEED-LLaMA repository.