BioVid: Autoregressive Video Generation with Biological Behavior Semantic Comprehension

BioVid introduces an autoregressive video generation framework that learns temporal structure from behavioral data rather than using fixed frame counts. The system uses a specialized tokenizer and transformer architecture to naturally determine when behavioral sequences end, matching real-world action duration distributions significantly better than existing methods.

BioVid addresses a fundamental limitation in current video generation systems: the treatment of video length as an external constraint rather than a data-driven property. Traditional frameworks impose predetermined frame counts or rely on text prompts, disconnecting generated content from how biological behaviors actually unfold in nature. This research demonstrates that action duration varies meaningfully across individuals and contexts, encoding information that systems should learn rather than override.

The technical approach combines two innovations. A Finite Scalar Quantization R3GAN tokenizer compresses video frames into discrete tokens while preventing the codebook collapse problem that undermines many quantization schemes. Building on this, a causal transformer then learns sequences autoregressively, learning to emit an End-of-Sequence token when behavioral completion occurs naturally. This design shifts the decision burden from human specification to emergent statistical patterns.

The experimental validation on human drinking behavior demonstrates substantial practical improvements. BioVid achieves a Wasserstein distance of 1.24 from ground truth length distributions, dramatically outperforming fixed-length baselines (6.05) and VideoGPT (15.48). This precision matters for applications requiring realistic behavioral simulation—from biomechanics research to embodied AI systems.

For the broader AI video generation landscape, BioVid signals movement toward more naturalistic, data-aligned generation paradigms. The framework's success suggests that encoding domain-specific structure—temporal in this case—produces both better fidelity and interpretability. Future work likely extends these principles to other behavioral domains where duration distributions carry semantic significance, making systems more aligned with how real-world phenomena unfold.

• →BioVid learns video duration distributions directly from data rather than imposing fixed frame counts or external constraints
• →FSQ-R3GAN tokenizer achieves high-fidelity frame encoding while eliminating codebook collapse problems
• →Causal transformer architecture learns to emit End-of-Sequence tokens based on behavioral semantic closure
• →Length distribution accuracy improves dramatically: 1.24 Wasserstein distance vs 6.05 for fixed-length and 15.48 for VideoGPT
• →Framework demonstrates domain-specific alignment yields better generalization in behavioral video synthesis