Stanford CME296 Diffusion & Large Vision Models | Spring 2026 | Lecture 7 - Evaluation

Short summary

Stanford graduate lecture covering evaluation methodologies for diffusion and large vision models, including Elo ratings, FID, CLIPScore, PickScore, LPIPS, and multimodal LLM-based judges. Spans reference-free metrics, perceptual similarity, and benchmarking approaches. Essential for researchers and practitioners evaluating vision model performance.

• •Comprehensive coverage of reference-free evaluation metrics (FID, CLIPScore, PickScore)
• •Perceptual similarity measures and multimodal LLM-based evaluation approaches
• •Practical frameworks for benchmarking and assessing vision model quality

Generated with AI, which can make mistakes.