Article

Researching living brain cells 4 years,TBC bringing neuron algorithms into Alibaba Wan 2.2, video generation is claimed to be up to 5 times faster

Beating AI News reports that a US company founded by two neurosurgeons TBC, has released the video model Mischief 1.

The new model is based on a modification of Alibaba's open-source Wan 2.2 model, adding less than 0.1%parameters.TBC claims that, at comparable image quality, video generation is at least 5 times faster, while inference costs are reduced by approximately 80%.

The technology comes from the team's continued 4 years of research on living neurons. Researchers cultured brain cells on a chip with 4096 electrodes, converted visual information into electrical stimulation, and recorded how neural activity propagated through space and decayed over time. Based on these patterns, they designed lightweight software modules and embedded them into the diffusion Transformer architecture of a video model, helping AI maintain consistency between frames without retraining the entire model from scratch.

Video models can be accelerated by reducing denoising steps, but reducing the number too much can make images blurry or distorted and make it harder to maintain consistency between frames.TBC previously conducted experiments using a AI model that generates Minecraft scenes in real time based on player actions. After adding the bio-inspired module, the model maintained stable visuals while reducing denoising steps. Combined with caching and sampling optimizations, the overall system's throughput increased by 4.4 times.

Now,TBC is taking this research path toward commercialization. Mischief 1 is now available through fal and OpenRouter, and currently supports generating videos of 3 seconds or 5 seconds from images. Generating a 5-second 480p video costs just 2.5 cents, while 720p costs 5 cents.

Original link https://m.theblockbeats.info/flash/374824