Reward Scheduling, Controllable Layouts, and Consumer-Grade 1080p Video: Today's Visual AI Digest

VideoCLIP-Score Is Broken: Why Your T2V Benchmarks Are Lying

The post exposes a recurring 3-15% over-count in CLIP-FID when video models are compared to image-adapted baselines, primarily due to 24-to-8 frame sampling mismatch. Researchers are already flagging older leaderboards as potentially misleading.