Visual AI Digest: Autoregressive Video Memory, Real-Time Diffusion, and the Consistency Arms Race

Research

Streaming Video Diffusion with Persistent Scene Memory

A new approach tackles the classic long-video problem: models forget what happened 5 seconds ago. This memory-bank trick keeps backgrounds and objects stable across minute-long generations without the usual drift, which is the missing piece for anything beyond short clips.

Identity-Locked Character Generation Across Poses and Scenes

Yet another attempt at solving the 'same character, different scene' problem that's plagued T2I since day one, this time using a lightweight adapter instead of full fine-tuning. If it generalizes, it's a big deal for anyone building comics, game assets, or branded content pipelines.

Tools

Few-Step Distillation Pushes T2I Latency Under 100ms

Another entry in the race to make diffusion feel instant — distilling a full multi-step sampler down to 2-4 steps while holding quality close to the teacher model. This is the kind of unglamorous engineering that actually determines whether these models ship in consumer apps versus staying research demos.

Training-Free Layout Control via Attention Rerouting

A neat trick that lets you dictate object placement in generated images just by manipulating attention maps at inference time — no fine-tuning, no ControlNet needed. Lowers the barrier for precise compositional control, which matters a lot for ad creative and product mockups.

Analysis

Benchmark Study: Where Text-to-Video Models Still Fail Physics

A systematic stress test across several leading T2V models finds they still botch basic physical interactions — liquids, collisions, gravity under occlusion. Useful reading if you're evaluating models for anything beyond aesthetic b-roll, since the gap between 'looks real' and 'behaves real' is still wide.