Test-Time Training cracks 1-minute AI cartoon generation

Most AI video models fall apart if you ask them to generate anything longer than five seconds. Characters morph into different creatures, rooms change layouts between cuts, and storytelling is practically out of the question.

Researcher Xiaolong Wang and collaborators unveiled a new approach that brings Test-Time Training (TTT) into Diffusion Transformers, successfully generating coherent, minute-long cartoon videos from text storyboards.

Turning hidden states into neural networks

Standard Transformers struggle with long videos because self-attention becomes computationally brutal over extended contexts. Cheaper alternatives like Mamba or sliding-window attention keep compute low, but their hidden states often lack the expressive power needed to track multi-scene narratives.

Wang's team tackled this using TTT modules. Instead of a traditional RNN hidden state, TTT treats the memory state as a machine learning model updated via gradient descent during inference.

By slotting TTT layers into a pre-trained 5B Transformer, the system produced a continuous one-minute Tom and Jerry cartoon. The prompt guided Jerry eating cheese, Tom teasing him, Jerry packing a suitcase for San Francisco, and Tom following his footprints to apologize with cheese in an alleyway.

In human evaluations across 100 generated videos per method, TTT beat baselines like Mamba 2, Gated DeltaNet, and sliding-window attention by 34 Elo points. While sliding-window attention duplicated Jerry and swapped kitchen furniture, TTT maintained character consistency across camera cuts.

Promising proof of concept, obvious artifacts

The project is still strictly a proof of concept, and the artifacts are easy to spot.

The authors noted that boxes morph across three-second video segments, cheese hovers unnaturally in mid-air instead of hitting the floor, and lighting in the kitchen suddenly spikes when Tom turns around. They attribute much of this to the relatively small 5B base model and unoptimized implementation, noting that resource constraints capped their test at 60 seconds.

Why it matters

Right now, AI video creators are essentially stitching together five-second clips in Premiere, praying characters don't randomly change clothes. If TTT scales, it could pave the way for true storyboard-to-video generation where a model maintains narrative continuity across complex multi-shot scripts.

It is not something you can drop into a production pipeline today. The physical glitches and lighting spikes show that 5B models still lack world understanding, meaning users shouldn't expect studio-grade cartoons just yet.

Still, we think this is one of the most clever architectural pivots in AI video. Instead of blindly brute-forcing Transformer context windows or settling for weak linear states, making memory an actively learning network at inference time could be the key to cracking long-form AI video.