Latent Space Video Synthesis With Temporal Consistency Preservation

via “video model training with temporal consistency objectives”

SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformer

Unique: Implements specialized temporal consistency losses (optical flow, perceptual) combined with block-causal attention masking during training, ensuring learned models maintain frame-to-frame coherence without post-processing

vs others: Achieves temporally coherent video generation through training-time consistency objectives rather than post-hoc smoothing, resulting in more natural motion and appearance transitions