As video generation shifts from specialized U-Net architectures to Diffusion Transformers (DiT), separating modalities is increasingly unnecessary. This session presents the Single-Stream paradigm, where text and video are embedded into a shared token space and processed as a single sequence by a standard PyTorch nn.TransformerEncoder, enabling joint attention across spatial, temporal, and semantic dimensions without modality-specific components.

We demonstrate Rectified Flow Matching in native PyTorch, replacing discrete noise schedules with straight-line probability paths parameterized by continuous flow time. The talk shows that multimodal DiT models reduce to conventional nn.TransformerEncoder layers applied to concatenated text and video tokens with modality-aware positional encodings.

Finally, we show how to optimize this architecture using torch.compile and FlashAttention (torch.nn.functional.scaled_dot_product_attention), producing a simpler, faster, and more maintainable training and inference pipeline. Attendees will leave with a deeper understanding of multimodal video generation using recent advances in AI and PyTorch.