uploads a photo of yourself, then records a voice message, and a digital version of “you” that blinks, opens your mouth, and has a natural expression is born. This is the currently very popular digital human technology.

With the widespread adoption of digital human live streaming, virtual employees, and AI short video tools, patent competition around core technologies has quietly begun.

Recently, the National Intellectual Property Administration announced typical cases of invalid patent reexamination for 2025, concluding the country’s first patent invalidation request in the virtual digital human field. The National Intellectual Property Administration decided to uphold the validity of the invention patent involved in the silicon-based intelligence.

This case focuses on the underlying algorithm patents for AI voice-driven image generation, clarifying for the first time the creative evaluation criteria for AI technology solutions in the digital human sector.

1. Basic facts of the case

• Patent holder: Nanjing Silicon-Based Intelligent Technology Group Co., Ltd

Focusing on deep research and application of artificial intelligence technology, its main business is providing customized “digital employees” services for B2B.

• Requester for invalidation declaration: Shenzhen Qutui Technology Co., Ltd

Mainly targeting consumer users, its digital human generation app can generate digital human broadcast videos with one click, forming a direct competition with silicon-based intelligence.
The patent in question: “Method, device, device, and storage medium for generating dynamic images based on audio,” was filed in January 2024.

2. Key

Points of Dispute The technical issue addressed by the patent in question is: how to generate a dynamic digital human highly identical to a specific person at low cost using a single photo and a segment of audio.

Conventional solutions require training deep learning networks with large amounts of high-quality video data, resulting in long training cycles and high costs.
The patent in question offers a new approach: users only need to input a photo of a target person and a reference audio clip, and the system can generate a talking, expressive, dynamic digital human.
Its core technical path is: first train the base model with general data, then personalize fine-tuning the model based on specific user photos, and finally use the adjusted model to generate video.
In other words, the focus of this patent innovation is not the result of “generating digital humans,” but the “personalized tuning” step added to the model training process.
However, among the invalid requests submitted by
Qutui Technology, the most critical comparative evidence is an existing technology called SadTalker.
SadTalker is a model that generates realistic synchronized videos using a single image and input audio. It can also achieve the effect of “input a photo + a segment of audio, output a video of conversation.”
On the surface, the two are indeed highly similar.
However, after in-depth analysis, the collegial panel discovered a fundamental difference.
SadTalker is a universal model. After training with sample set data, it can directly generate videos based on any user’s input images and audio, without re-tuning the model after training. Its architecture is an end-to-end “training-to-apply” model.
The patent involved is completely different. It is based on an already trained generative network model, using user-input reference images to personalize the model before generating videos. Its architecture is “Training-Personalization-Application.”

The panel determined that the difference lies in the model training methods. The patent in question solves the technical problem of “making generated results more tailored to specific users while reducing training costs” by adding a “personalized adjustment” step.
Qutui Technology also advocated combining Evidence 1 (SadTalker) with Evidence 2 (a 3D facial reconstruction method), arguing that the patent involved was obvious.
The collegial panel also did not support it.
Although Evidence 2 also involves the model training phase, its disclosed content is: generating reconstructed images based on training images, using mixed loss functions to evaluate differences between reconstructed and training images, and performing backpropagation to train 3D facial reconstruction models.
However, the panel pointed out: Evidence 2 does not involve digital human generation, and its training process is completed in one go. Subsequently, the trained model is directly used to reconstruct 3D faces from any user photo, without any minor model adjustments.
Evidence 2 can serve as a replacement for a submodel in Evidence 1, but does not provide insights for personalized customization of the overall model.
Regarding this case, Associate Professor Hu Jingjing from the School of Civil and Commercial Law at Southwest University of Political Science and Law commented: “It cannot be done solely because it belongs to the field of virtual digital human generation; the fine-tuning and improvement methods for the presumed training stage are obvious.” Existing technologies generally adopt an end-to-end ‘training-application’ architecture, without disclosing technical methods for ‘using homogeneous reference images for targeted secondary tuning of pre-trained models,’ nor has it developed an improvement approach for ‘low-cost generation of high-precision dynamic images of specific people.’ ”

3. Case Demonstration Significance

The value of this case goes beyond individual victories, providing important guidance for patent examination and patent applications in the entire AI industry.

First, it establishes a methodology for determining the inventiveness of AI patents.
The National Intellectual Property Administration pointed out that technical solutions involving AI applications typically include three stages: model design, model training, and model application. During the review, the technical solution should be accurately judged from four dimensions: processing purpose, treatment object, processing result, and result usage, thereby determining the substantive difference from prior art.
In an interview, examiner Wu Shifen emphasized that the focus of the dispute lies in “how to accurately understand the technical solution of the claims, and whether the prior art provides insights for improving model training methods.”

Second, substantial improvements during the training phase are key to innovation.
The panel clearly stated: “The model training stage is a specific and important part of the technical solutions involved in AI applications.” ”
If the difference between claims and prior art lies in the model training method, and the training method delivers more precise or lower-cost technical results, and the prior art overall does not provide improvement insights, then this distinction is sufficient to support the determination of inventiveness.
Specifically in this case, the collegial panel determined that, based on the characteristics of the digital human industry, the patent in question achieved technical results with lower model costs and more user-friendly generation results by improving model training methods.

Third, it points out the direction for drafting AI patent applications.
Associate Professor Hu Jingjing pointed out: “The inventiveness review of training inventions in the field of artificial intelligence needs to focus on differentiated substantial improvements in algorithms during the training phase, such as parameter tuning and data processing workflows, and the corresponding effects they produce.” ”
For applicants, the writing of patent application documents should focus on the technical problems to be solved during the training phase, fully recording the technical features such as the training method steps, parameter optimization logic, and training sample selection rules, clearly explaining the technical effects brought by this method, and highlighting the substantial improvements that distinguish it from general model training.

4. Conclusion

This first domestic digital human patent invalidation case sends a clear signal: in the AI field, substantial innovation in model training methods can form a strong patent barrier. For companies in this sector, it’s not just about speed in technology R&D, but also about the depth of patent layout.

As competition in the digital human industry intensifies, the threshold for replicating product functions is getting lower. To maintain innovation achievements in competition, companies must not only accelerate technology R&D iteration but also excel in patent mining and layout for underlying algorithms and training processes, using intellectual property to safeguard technological innovation.

MeritsIP: Your Global Intellectual Property Partner

At MeritsIP, we provide comprehensive IP services in biomedicine, medical devices, manufacturing, semiconductors, and AI. We support 200+ top clients globally with strategic, high-quality IP services.

Stay Connected

• • Website: https://www.meritsipglobal.com

• • LinkedIn: Follow us on LinkedIn

• • Email: meritsip@meritsandtree.com

© 2025 MeritsIP. All Rights Reserved.