April 2026 Special Course - Optional Lecture 1: Datasets, Captions, and Data Filtering

Why it matters:
Model quality is not just architecture. Data quality, coverage, and filtering often matter just as much.

Topics:

• Where image-text and video-text data come from

• Caption generation

• Deduplication

• NSFW and copyright filtering

• Why video data is harder than image data

• Why hands, text rendering, and rare actions often fail

Suggested materials:

• LAION

• DataComp

• WebVid-10M

• HD-VILA

• Data sections of SVD, Imagen Video, and HunyuanVideo reports