Custom Text-to-Video Adaptation Without Custom Video Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The generation of customized videos is challenging due to the scarcity of customized video data, leading to inefficiencies and inaccuracies when integrating text-to-image models into text-to-video frameworks, and the computational load is high due to the mismatch in feature distribution between these models.

Innovation Solution

A framework that modifies a text-to-video model with motion and spatial adapter blocks, trained using natural and custom motion-free videos, to align feature distributions and control motion, allowing customization without extensive video data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If text-to-image models are integrated into text-to-video frameworks to enable customization, then customization capability is improved, but computational load increases and feature distribution mismatch occurs

Engineering Contradiction:
Improvecustomization capabilityVSAvoidcomputational load
Core Design Contradiction:
Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The model is segmented into distinct components: a frozen text-to-image model for spatial feature extraction, a text encoder for condition processing, and a video generation model with trainable adapter modules. This segmentation allows the heavy T2I model to remain frozen while only lightweight components are trained, reducing computational load during customization while maintaining adaptability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

An adapter module serves as an intermediary between the frozen T2I model and the video generation model. This adapter translates the feature distribution from the T2I model into a format compatible with the video model, resolving the feature distribution mismatch without requiring full integration or retraining of the T2I model, thus reducing computational overhead.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If text-to-image models are integrated into text-to-video frameworks with direct weight replacement, then customization is achieved, but feature distribution mismatch degrades video generation quality

Engineering Contradiction:
Improvecustomization capabilityVSAvoidvideo generation quality
Core Design Contradiction:
Adaptability or versatilityVSManufacturing precision

Solution Approach 1:

An adapter module is introduced as an intermediary component between the frozen T2I model and the video generation model. This adapter learns to transform the feature distribution from the T2I model into a format compatible with the video model, preventing feature distribution mismatch and maintaining high video generation quality while enabling customization.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

Instead of replacing model weights directly, the system changes parameters by training only the adapter module and text encoder while keeping the T2I model frozen. This parameter-efficient approach adjusts the feature transformation process to match distributions without altering the core T2I model weights, preserving video quality while enabling customization.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If extensive customized video datasets are collected for training, then customization accuracy is improved, but data availability decreases and collection complexity increases

Engineering Contradiction:
Improvecustomization accuracyVSAvoiddata availability
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The text-to-image model is pre-trained on extensive image datasets before being frozen and reused for video generation tasks. This preliminary training on abundant image data allows the model to learn robust spatial features that can be transferred to video generation without requiring extensive customized video datasets, effectively bridging the data availability gap.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system copies and reuses the pre-trained T2I model weights for the video generation task. Instead of training a new model from scratch on scarce video data, the approach copies the learned spatial representations from the T2I model and adapts them to video generation through the adapter module, achieving high customization accuracy without extensive video data collection.

Inventive Principle:
Principle #26Copying

4Adaptability or versatility

If the T2V model is trained to adapt to T2I model features, then customization capability is improved, but training time and computational resources increase

Engineering Contradiction:
Improvecustomization capabilityVSAvoidtraining time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The training process is segmented into two phases: (1) freezing the T2I model and training only the adapter and text encoder on image data, and (2) fine-tuning the video generation components. This segmentation reduces training time by avoiding full-model retraining while still achieving adaptation to T2I features through the lightweight adapter module.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system changes parameters by training only a small subset of parameters (adapter module and text encoder) rather than all parameters in the T2V model. This parameter-efficient fine-tuning approach enables the model to adapt to T2I features with minimal training time and computational resources while maintaining customization capability.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4668267A1Customized video generation
Publication Date: 2025.12.24 GOOGLE LLC
  • EP4668267A1 patent drawingFigure 1
  • EP4668267A1 patent drawingFigure 2
  • EP4668267A1 patent drawingFigure 3

AI summary

Provided are systems and methods for generating custom text-to-video (T2V) models starting from a custom text-to-image (T2I) model and without requiring customized video data. The proposed techniques can be particularly beneficial for applications where video data of a specific subject or style is not available. For example, the proposed approach can be used to create custom videos from a small set of custom still images or generate videos in a specific custom artistic style without having prior videos in that style.