Sora Gets Competition, ControlNet Goes Viral Again, and New Benchmarks Shake Up Text-to-Video Rankings

New Video Generation Benchmark Exposes Gaps Between Closed and Open-Source Models on Motion Coherence

Researchers introduce a structured evaluation framework that stress-tests temporal consistency and physics plausibility — areas where many popular models quietly underperform. The results suggest current leaderboard rankings based on FID alone are misleading practitioners.

T2I Personalization Paper Achieves Subject Consistency Without Test-Time Fine-Tuning

A new approach achieves strong subject fidelity in text-to-image generation at inference time only, skipping the costly per-subject training loop that makes personalization impractical at scale. This could meaningfully change how e-commerce and ad-tech teams approach product image generation.