🗣 Seminar Talks

Invited talk at Salim Habib University covering the progression from diffusion-based generative modeling to reasoning- and agent-driven computer vision systems. The presentation introduces diffusion fundamentals, restoration-oriented diffusion pipelines, reasoning-based restoration methods such as Chain-of-Zoom and Chain-of-Restoration, and agentic workflows including RestoreAgent and 4KAgent.

Topics Covered

• Diffusion models for generative modeling and image restoration, including StableSR and efficient fine-tuning.
• Reasoning-based restoration strategies such as Chain-of-Zoom and Chain-of-Restoration.
• Agentic restoration systems including RestoreAgent and 4KAgent, with planning, execution, reflection, and tool use.
• Foundation-model examples spanning segmentation, self-supervised learning, and world models.

Seminar covering the use of large language models and vision-language approaches in visual reasoning and restoration. We explored recent innovations in step-wise reasoning and multimodal agents for complex image tasks.

Referenced Papers

• Improving Language Model Reasoning with Self-Motivated Learning
• RestoreAgent: Autonomous Image Restoration Agent via Multimodal LLMs
• Chain-of-Restoration: Universal Step-by-Step Restorers

Presents a multi-task learning viewpoint on modern diffusion models. It discusses task transferability, negative transfer challenges, and the use of mixture-of-experts to solve denoising in a synergistic fashion.

Referenced Papers

• Addressing Negative Transfer in Diffusion Models
• Switch Diffusion Transformer

A deep dive into diffusion-based models tailored for specific restoration tasks, including Ambient Diffusion, ResShift for super-resolution, and DiffBIR's two-stage blind restoration pipeline.

Referenced Papers

• Ambient Diffusion
• ResShift: Efficient SR via Residual Shifting
• DiffBIR: Diffusion-based Blind Image Restoration

Highlights from two AAAI-24 papers: Any-Size-Diffusion (ASD), a two-stage pipeline for text-to-image synthesis at arbitrary resolutions, and ResDiff, which unites CNN priors with diffusion guidance for efficient single-image super-resolution.

Referenced Papers

• Any-Size-Diffusion: Text-Driven Synthesis for Any-Size HD Images
• ResDiff: Combining CNN and Diffusion for Image SR

Summary session from NeurIPS 2024, with analyses of several key papers presented at the conference.