Memory, Physics, and Thinking: Sharpening Visual Consistency

Paper

Memory-Locked Tokenizer Solves the Temporal Drift in Long Video

This addresses a major bottleneck in video generation: consistency across long clips. By using an adaptive memory system, it prevents visual drift and flickering, which is crucial for anyone trying to produce watchable AI video.

Embracing Chaos: Physics-Aware Rewards Guide Video Diffusion

Instead of fighting the diffusion process, this approach flips the script and intentionally brings objects into collision. It’s a fascinating way to handle complex multi-object interactions that typically confuse standard generators.

Slow Thinking for Vision: ReasoVLM Boosts Instruction Following

Thinking models are coming to the visual domain to handle complex compositions. This is a major step toward fixing the 'two left hands' problem by using step-by-step reasoning for spatial relationships.

Tools

CogVideoX 1.5: The Open-Source Standard Bumps Up Coherence and Fidelity

CogVideoX was already a strong open-source player, and this update tightens the coherence and aesthetic quality. It's the practical choice right now for developers wanting a deployable, high-quality model that moves beyond basic demos.