Diffusion Transformer Video Generation With Segment Memory Vectors

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional systems struggle to generate long temporal sequences, such as videos, with high quality and coherence due to computational constraints and error accumulation, especially when conditioned on a small number of previous frames, leading to rapid quality degradation.

Innovation Solution

A latent denoising neural network is used to generate long temporal sequences by conditioning on a memory vector representing preceding segments, employing a diffusion Transformer architecture that updates latent representations efficiently, reducing computational resources and maintaining quality over time.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Duration of action of moving object

If conventional diffusion models are used to generate videos autoregressively, then video generation is possible, but computational resources and memory requirements increase significantly with video length

Engineering Contradiction:
Improvevideo lengthVSAvoidcomputational resources
Core Design Contradiction:
Duration of action of moving objectVSUse of energy by moving object

Solution Approach 1:

The video generation process is divided into multiple segments, where each segment is generated independently conditioned on a compressed memory representation of previous segments. This segmentation allows the model to handle long videos without processing the entire sequence at once, reducing computational resource requirements while maintaining long-term consistency.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extracts and utilizes only the essential temporal information from previous video segments by compressing it into a memory vector. This extraction process removes redundant data while preserving the key temporal context needed for generating subsequent segments, thereby reducing memory and computational requirements.

Inventive Principle:
Principle #2Taking out (Extraction)

2Reliability

If a large window size of previous video frames is used for conditioning, then long-term consistency is improved, but computational constraints are violated

Engineering Contradiction:
Improvelong-term consistencyVSAvoidcomputational constraints
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent transforms the conditioning approach by changing the parameter representation from raw video frames to a compressed memory vector. This parameter transformation maintains the essential temporal information needed for long-term consistency while dramatically reducing the computational complexity associated with processing large numbers of high-dimensional video frames.

Inventive Principle:
Principle #35Parameter changes

3Productivity

If autoregressive generation is used with limited window size, then computational constraints are satisfied, but error accumulation causes rapid quality drop

Engineering Contradiction:
Improvegeneration efficiencyVSAvoidvideo quality
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent introduces a memory vector as an intermediary between previous video segments and the current segment being generated. This intermediary compresses and preserves the essential temporal context from previous segments, allowing the model to maintain high video quality over long durations without suffering from error accumulation that plagues systems using limited window sizes.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4654079A1Generating temporal sequences using diffusion transformer neural networks
Publication Date: 2025.11.26 GOOGLE LLC
  • EP4654079A1 patent drawingFigure 1A
  • EP4654079A1 patent drawingFigure 1B
  • EP4654079A1 patent drawingFigure 2A

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating an output temporal sequence of data elements conditioned on an input. In one aspect, a method comprises: obtaining the input, wherein the input comprises a noise input comprising a respective latent representation for each of a plurality of segments of the temporal sequence; updating, for each segment, the latent representation for the segment using a latent denoising neural network, the updating comprising, for each segment other than the first segment: obtaining a memory vector representing one or more hidden states generated by the latent denoising neural network when updating the latent representations for one or more preceding segments; updating the latent representation for the segment at each of a plurality of iterations; and generating the output temporal sequence of data elements by processing the latent representations for the plurality of segments.