Even More MASSIVE Video AI Upgrades & New Models!!! it just does not stop!

Spatio-temporal skip guidance (STG) is emerging as a practical upgrade that makes video diffusion models produce sharper details and more consistent motion across frames—without abandoning the familiar classifier-free guidance workflow. In side-by-side demos, smoke looks more physically grounded, facial features sharpen instead of turning into blobs, and motion stays coherent rather than “morphing” between frames. The most striking claim is temporal scaling: guidance that works for single images is one thing, but STG is presented as a way to keep improvements stable across an entire clip, including hair, lighting highlights, and background structures like trees, clouds, and castles.

STG is positioned as a “no-compromises” add-on that can run on its own or alongside standard classifier-free guidance. Examples highlight the difference: a butterfly that’s blurry and mushy in the baseline becomes crisp enough to show individual legs and antennae; a woman’s lips and teeth form more clearly instead of collapsing into random skin-like shapes; and a 3D cat’s viewpoint stays consistent during a jump rather than flipping positions abruptly. The workflow is also framed as accessible for open-source video generators, with the demos attributed to the Mochi generator using STG enabled. The same technique is said to be applicable to other systems such as Stable Video Diffusion and Open Sora, with the promise of improvements “across the board.”

The broader theme is that video quality is accelerating through both algorithmic guidance and new model releases. A major new fully open-source video generation model from 10cent is described as having strong physics and lighting understanding, with open weights and code. The tradeoff is compute: early users reportedly need around 60 GB of VRAM, far beyond typical consumer GPUs. Still, the open-source framing is treated as a community win—lowering the barrier for experimentation and potentially pushing hardware requirements down over time.

The 10cent demos emphasize complex, multi-object scenes: consistent chair patterns while people move in the background, slow-motion fire lighting a child’s face realistically, camels walking with relatively stable feet, and sand behaving plausibly as it falls from a person’s hands. Even a cat eating a cheeseburger is cited as a difficult test for leg coordination and multi-part object continuity, with the model described as getting “pretty darn well” despite minor glitches.

Alongside model upgrades, new tools target specific styles. Minimax (Hailu AI) introduces 12 vo1 live, aimed at transforming 2D illustration art into more dynamic motion—especially anime-like sequences—while maintaining character consistency. The tool is explicitly not presented as a system that takes an existing video of someone speaking and transfers that motion; instead, it’s trained for illustration-driven animation, with demos showing hair moving independently and frame-hold behavior that resembles traditional animation timing.

Finally, Google’s VO AI is mentioned as available via Vertex AI in private preview, with early impressions described as less impressive than newer top-tier models—suggesting it may have lost momentum since its announcement. The segment closes with speculation that OpenAI could be preparing another release soon, with community chatter centered on Sora and other OpenAI developments, though no concrete confirmation is provided.