Granted Patent B2

US 12,456,245 · App. 18/478,638 · Granted Oct 28, 2025

Inventor: Hau Nghiep Phan (Montreal, CA)

Assignee: Electronic Arts Inc.

G06T13/80G06T3/4053G06T19/20G06T2219/2024

Abstract

Claims (80)

1. A computer-implemented method comprising:

accessing a first set of machine learning models trained to generate, via a latent variable space, a three-dimensional (3D) mesh and a plurality of two-dimensional (2D) texture maps corresponding to the three-dimensional mesh;

obtaining a first set of input information configured to generate a first facial model having a first identity;

generating, by the first set of machine learning models, the first facial model from the first set of input information, wherein the first facial model includes a plurality of 2D texture maps and a 3D facial mesh of a first face;

accessing a second machine learning model trained to generate photorealistic 2D images based on an input image; and

generating, using the second machine learning model, at least one target photorealistic 2D image based on the first facial model;

accessing a differentiable rendering engine, wherein the differentiable rendering engine is configured to modify facial models based on target 2D images;

modifying, by the differentiable rendering engine, the first facial model based on the at least one target photorealistic 2D image; and

outputting, by the differentiable rendering engine, an enhanced first facial model, wherein the enhanced first facial model includes a modified version of the plurality of 2D texture maps and a modified version of the 3D facial mesh.

2. The computer-implemented method of claim 1 , further comprising, prior to generating the at least one target photorealistic 2D image:

accessing a diffusion-based machine learning model configured to increase resolution of a 2D image;

generating, using the diffusion-based machine learning model, an upscaled 2D texture map for each of the plurality of 2D texture maps, wherein a resolution of each of the plurality of 2D texture maps is increased to a higher resolution.

3. The computer-implemented method of claim 1 , wherein modifying the first facial model comprises:

generating a plurality of input images from character facial model based on the plurality of 2D texture maps and 3D facial mesh of the first face from a plurality of different angles;

generating, by the second machine learning model, a plurality of target photorealistic 2D images from the plurality of input images; and

extracting, by the differentiable rendering engine, a plurality of 3D surface properties from the plurality of photorealistic images;

adjusting the plurality of 2D texture maps and 3D facial meshes of the first face based on the plurality of 3D surface properties.

4. The computer-implemented method of claim 3 , wherein the plurality of 2D texture maps comprise a diffuse texture map and a normal texture map generated by the first set of machine learning models, and a roughness texture map generated by an independent process.

5. The computer-implemented method of claim 3 , further comprises iteratively adjusting the 2D texture maps and the 3D facial meshes until a difference threshold is satisfied.

6. The computer-implemented method of claim 5 , further comprising:

determining a difference between facial features of at least one 2D render of the adjusted facial model and at least one target photorealistic image; and

determining whether the difference satisfies the difference threshold.

7. The computer-implemented method of claim 1 , wherein the first facial model is generating having a first expression, and the first set of machine learning models is configured to generate the first facial model for a plurality of expressions.

8. The computer-implemented method of claim 6 , wherein the plurality of 2D texture maps comprise at least a diffuse map and a normal map for each expression of the plurality of expressions.

9. The computer-implemented method of claim 1 , wherein the first set of machine learning models comprise:

an identity machine learning model trained to generate identity information;

an expression machine learning model trained to generate expression information;

a texture map machine learning model trained to generate 2D texture maps; and

a mesh machine learning model trained to generate 3D meshes;

wherein generating the first facial model comprises:

generating, by the identity machine learning model, identity information representative of an invariant identity of the first face;

generating, by the expression machine learning model, expression information from the first set of input information, wherein the expression information includes a defined set of expressions of the first face;

generating, by the texture map machine learning model, the 2D texture maps of the first face in the defined set of expressions from the identity information and the expression information; and

generating, by the mesh machine learning model, the 3D facial meshes of the first face in the defined set of expressions from the identity information and the expression information.

10. A system comprising one or more computers and non-transitory computer storage media storing instructions that when executed by the one or more computers, cause the one or more computers to perform operations comprising:

accessing a first set of machine learning models trained to generate, via a latent variable space, a three-dimensional (3D) mesh and a plurality of two-dimensional (2D) texture maps corresponding to the three-dimensional mesh;

obtaining a first set of input information configured to generate a first facial model having a first identity;

generating, by the first set of machine learning models, the first facial model from the first set of input information, wherein the first facial model includes a plurality of 2D texture maps and a 3D facial mesh of a first face;

accessing a second machine learning model trained to generate photorealistic 2D images based on an input image; and

generating, using the second machine learning model, at least one target photorealistic 2D image based on the first facial model;

accessing a differentiable rendering engine, wherein the differentiable rendering engine is configured to modify facial models based on target 2D images;

modifying, by the differentiable rendering engine, the first facial model based on the at least one target photorealistic 2D image; and

outputting, by the differentiable rendering engine, an enhanced first facial model, wherein the enhanced first facial model includes a modified version of the plurality of 2D texture maps and a modified version of the 3D facial mesh.

11. The system of claim 10 , wherein the instructions further configure the one or more computers to perform operations comprising, prior to generating the at least one target photorealistic 2D image:

accessing a diffusion-based machine learning model configured to increase resolution of a 2D image;

generating, using the diffusion-based machine learning model, an upscaled 2D texture map for each of the plurality of 2D texture maps, wherein a resolution of each of the plurality of 2D texture maps is increased to a higher resolution.

12. The system of claim 10 , wherein the instructions further configure the one or more computers to perform operations when modifying the first facial model comprising:

generating a plurality of input images from character facial model based on the plurality of 2D texture maps and 3D facial mesh of the first face from a plurality of different angles;

generating, by the second machine learning model, a plurality of target photorealistic 2D images from the plurality of input images; and

extracting, by the differentiable rendering engine, a plurality of 3D surface properties from the plurality of photorealistic images;

adjusting the plurality of 2D texture maps and 3D facial meshes of the first face based on the plurality of 3D surface properties.

13. The system of claim 12 , wherein the plurality of 2D texture maps comprise a diffuse texture map and a normal texture map generated by the first set of machine learning models, and a roughness texture map generated by an independent process.

14. The system of claim 12 , wherein the instructions further configure the one or more computers to perform operations comprising iteratively adjusting the 2D texture maps and the 3D facial meshes until a difference threshold is satisfied.

15. The system of claim 14 , wherein the instructions further configure the one or more computers to perform operations comprising:

determining a difference between facial features of at least one 2D render of the adjusted facial model and at least one target photorealistic image; and

determining whether the difference satisfies the difference threshold.

16. The system of claim 10 , wherein the first facial model is generating having a first expression, and the first set of machine learning models is configured to generate the first facial model for a plurality of expressions.

17. The system of claim 16 , wherein the plurality of 2D texture maps comprise at least a diffuse map and a normal map for each expression of the plurality of expressions.

18. The system of claim 10 , wherein the first set of machine learning models comprise:

an identity machine learning model trained to generate identity information;

an expression machine learning model trained to generate expression information;

a texture map machine learning model trained to generate 2D texture maps; and

a mesh machine learning model trained to generate 3D meshes;

wherein generating the first facial model comprises:

generating, by the identity machine learning model, identity information representative of an invariant identity of the first face;

generating, by the expression machine learning model, expression information from first set of input information, wherein the expression information includes a defined set of expressions of the first face;

generating, by the texture map machine learning model, the 2D texture maps of the first face in the defined set of expressions from the identity information and the expression information; and

generating, by the mesh machine learning model, the 3D facial meshes of the first face in the defined set of expressions from the identity information and the expression information.

19. Non-transitory computer-readable medium storing computer-executable instructions that when executed by a system of one or more computers, cause the one or more computers to perform operations comprising:

accessing a first set of machine learning models trained to generate, via a latent variable space, a three-dimensional (3D) mesh and a plurality of two-dimensional (2D) texture maps corresponding to the three-dimensional mesh;

obtaining a first set of input information configured to generate a first facial model having a first identity;

generating, by the first set of machine learning models, the first facial model from the first set of input information, wherein the first facial model includes a plurality of 2D texture maps and a 3D facial mesh of a first face;

accessing a second machine learning model trained to generate photorealistic 2D images based on an input image; and

generating, using the second machine learning model, at least one target photorealistic 2D image based on the first facial model;

accessing a differentiable rendering engine, wherein the differentiable rendering engine is configured to modify facial models based on target 2D images;

modifying, by the differentiable rendering engine, the first facial model based on the at least one target photorealistic 2D image; and

outputting, by the differentiable rendering engine, an enhanced first facial model, wherein the enhanced first facial model includes a modified version of the plurality of 2D texture maps and a modified version of the 3D facial mesh.

20. The non-transitory computer-readable medium of claim 19 , wherein, prior to generating the at least one target photorealistic 2D image, the instructions further configure the one or more computers to perform operations comprising:

accessing a diffusion-based machine learning model configured to increase resolution of a 2D image;

generating, using the diffusion-based machine learning model, an upscaled 2D texture map for each of the plurality of 2D texture maps, wherein a resolution of each of the plurality of 2D texture maps is increased to a higher resolution.

Assignments (1)

Continuity (1)

References Cited (296)