August 4, 2025 10:00 AM (GMT+7) → 12:20 PM

<aside> ℹ️

WS note: Thầy Triết

0> Multimedia system

New Generation of Multimedia

1> Image Generation and Editing

Text - to - Image:

Prompt to generate image

• In space n- dims, point, vector in this space reflect to other space (encode - decode)

Latent Diffusion Model

• PP: “High-Resolution Image Synthesis with Latent Diffusion model”, CVPR2022

Zero-shot image editing

• Photoshop: cut, past, color, editing
• compare to AI: generating new image,
• use cross-attention to keep some features
• Ex: (Qualitative results), thematic, Intelligent concept design system, creative image editing… human+AI to edit image(pose transition, obj replacement-fixed, dynamic, )

2> More

Personalized text to image: Keep image features, change..

Trade-off between reconstruction and editability:

Concept flows: Learning but not effect or noise with previous

**KronA-WED Adapter:

Learning Strategy: from original image, generate image with different scenes…

Potential apps: adsvertersing…

3> LLM-Empowered Generation

Automated image recognition framework:

• Prompt crafting in AIR-Gen+Textual prompt Extracting in AIR-Aug ⇒ ImageGeneration ⇒ Duplicated and outlier removal ⇒ output
• Ex: Forest, Early forest fire…
• Concept interpretation
• CLIP usually using for encode data

4>

• Question: Safety, Fraud, Valid…
• Interact system and allow it generate image,… must be careful
• Limitation: abnormal image not like real image… improve qualitive results: powerful hardware, math,… prompt to generate best image
• Generic if data big enough, so should use specific data that service your purpose if not enough data
• Difficult to evaluate results, solutions: compare results if assign work for multiple agents and random check if all agents have the same results…

</aside>