Editing digital images utilizing a neural network with an in-network rendering layer
The present disclosure includes methods and systems for generating modified digital images utilizing a neural network that includes a rendering layer. In particular, the disclosed systems and methods can train a neural network to decompose an input digital image into intrinsic physical properties (e.g., such as material, illumination, and shape). Moreover, the systems and methods can substitute one of the intrinsic physical properties for a target property (e.g., a modified material, illumination, or shape). The systems and methods can utilize a rendering layer trained to synthesize a digital image to generate a modified digital image based on the target property and the remaining (unsubstituted) intrinsic physical properties. Systems and methods can increase the accuracy of modified digital images by generating modified digital images that realistically reflect a confluence of intrinsic physical properties of an input digital image and target (i.e., modified) properties.
1. A system for generating modified digital images from input digital images, comprising:
one or more memories, comprising:
a neural network comprising a rendering layer trained to generate synthesized digital images from input digital images portraying diffuse materials and input digital images portraying specular materials; and
an input digital image; and
at least one computing device storing instructions thereon, that, when executed by the at least one computing device, cause the system to:
predict, utilizing the neural network, a material property set, a surface orientation map, and an illumination environment map based on the input digital image;
replace at least one of the material property set, the surface orientation map, or the illumination environment map with a target material property set, a target surface orientation map, or a target illumination map; and
utilize the rendering layer of the neural network to generate a modified digital image from the input digital image based on at least one of the target material property set, the target surface orientation map, or the target illumination map and at least two of the material property set, the surface orientation map, or the illumination environment map.
2. The system of claim 1 , further comprising instructions that, when executed by the at least one computing device, cause the system to:
replace the material property set with the target material property set; and
utilize the rendering layer of the neural network to generate the modified digital image based on the target material property, the surface orientation map, and the illumination environment map.
3. The system of claim 1 , wherein the input digital image portrays a first material corresponding to the material property set and portrays a second material, and further comprising instructions that, when executed by the at least one computing device, further cause the system to:
predict, utilizing the neural network, the material property set corresponding to the first material and a second material property set corresponding to the second material;
replace the material property set with the target material property set;
replace the second material property set with a second target material property set; and
utilize the rendering layer of the neural network to generate the modified digital image based on the target material property set, the second target material property set, the target surface orientation map, and the target illumination map.
4. The system of claim 1 , further comprising instructions that, when executed by the at least one computing device, cause the system to predict, utilizing the neural network, the material property set based on a parametric representation that is differentiable with respect to incoming light directions, outgoing light directions, and material properties.
5. The system of claim 1 , further comprising instructions that, when executed by the at least one computing device, cause the system to predict, utilizing the neural network, the surface orientation map, wherein the surface orientation map comprises an RGB digital image having a plurality of pixels with RBG values, where the RBG value of each pixel of the plurality of pixel encodes x, y, and z dimensions of a surface orientation of each pixel of an object portrayed in the digital image.
6. The system of claim 1 , further comprising instructions that, when executed by the at least one computing device, cause the system to generate the modified digital image such that the modified digital image portrays a specular material.
7. A non-transitory computer readable medium storing instructions thereon that, when executed by at least one processor, cause a computer system to:
receive an input digital image portraying an object within an illumination environment, wherein the object comprises a surface of a material, and wherein the surface has a surface normal direction;
predict, utilizing a neural network comprising a rendering layer trained to generate synthesized digital images from input digital images portraying diffuse materials and input digital images portraying specular materials, a material property set, a surface orientation map, and an illumination environment map based on the input digital image;
replace at least one of the material property set, the surface orientation map, or the illumination environment map with a target material property set, a target surface orientation map, or a target illumination map; and
utilize the rendering layer of the neural network to generate a modified digital image from the input digital image based on at least one of the target material property set, the target surface orientation map, or the target illumination map and at least two of the material property set, the surface orientation map, or the illumination environment map.
8. The non-transitory computer readable medium of claim 7 , further comprising instructions that, when executed by the at least one processor, cause the computer system to:
replace the at least one of the material property set, the surface orientation map, or the illumination environment map by replacing the material property set with the target material property set; and
utilize the rendering layer of the neural network to generate the modified digital image by generating the modified digital image based on the target material property set, the surface orientation map, and the illumination environment map.
9. The non-transitory computer readable medium of claim 7 , further comprising instructions that, when executed by the at least one processor, cause the computer system to:
replace the at least one of the material property set, the surface orientation map, or the illumination environment map by replacing the illumination environment map with the target illumination map; and
utilize the rendering layer of the neural network to generate the modified digital image by generating the modified digital image based on the material property set, the surface orientation map, and the target illumination map.
10. The non-transitory computer readable medium of claim 7 , further comprising instructions that, when executed by the at least one processor, cause the computer system to:
receive a request to modify the surface normal direction by replacing the object with a second object that comprises a second surface of a second material having a second normal direction; and
wherein the modified digital image portrays the second normal direction of the second object with the material of the object in the illumination environment of the input digital image.
11. The non-transitory computer readable medium of claim 7 , further comprising instructions that, when executed by the at least one processor, cause the computer system to:
replace the at least one of the material property set, the surface orientation map, or the illumination environment map by replacing the surface orientation map with the target surface orientation map; and
utilize the rendering layer of the neural network to generate the modified digital image by generating the modified digital image based on the material property set, the target surface orientation map, and the illumination environment map.
12. The non-transitory computer readable medium claim 7 , further comprising instructions that, when executed by the at least one processor, cause the computer system to generate the modified digital image such that the modified digital image portrays a specular material.
13. The non-transitory computer readable medium of claim 7 , further comprising instructions that, when executed by the at least one processor, cause the computer system to predict, utilizing the neural network, the material property set based on a parametric representation that is differentiable with respect to incoming light directions, outgoing light directions, and material properties.
14. In a digital medium environment for computer-based editing of digital images, a computer-implemented method of generating a trained neural network for rendering modified digital images, comprising:
providing a training input image to a neural network, wherein the training input image corresponds to a training material property set, a training surface orientation map, and a training illumination environment map;
predicting, via the neural network, a material property set, a surface orientation map, and an illumination environment map based on the training input image;
generating a synthesized training image, via a rendering layer of the neural network, based on the predicted material property set, the predicted surface orientation map, and the predicted illumination environment map;
determining a loss measure by comparing:
the synthesized training image with the training input image,
the training material property set with the predicted material property set, and
the training surface orientation map with the predicted surface orientation map; and
training the neural network utilizing the determined loss measure to generate a modified digital image from an input digital image.
15. The method of claim 14 , further comprising:
accessing a training target image and a training target property, wherein the training target image comprises the training input image modified to portray the training target property:
providing the training target property to the neural network; and
generating a synthesized target image, via the rendering layer of the neural network, based on the training target property and at least two of the following: the predicted material property set, the predicted surface orientation map, and the predicted illumination environment map.
16. The method of claim 15 , wherein determining the loss measure further comprises comparing the training target image with the synthesized target image.
17. The method of claim 15 , wherein the training target property comprises at least one of a target material property set different than the training material property set, a target training surface orientation map different than the surface orientation map, or a target illumination environment map different from the training illumination environment map.
18. The method of claim 15 , wherein predicting the material property set based on the training input image comprises utilizing a parametric representation that is differentiable with respect to incoming light directions, outgoing light directions, and material properties.
19. The method of claim 18 , wherein the parametric representation comprises a linear combination of a bidirectional reflectance distribution function database comprising bidirectional reflectance distribution functions corresponding to a plurality of materials.
20. The method of claim 15 , wherein the illumination environment map comprises a spherical panorama digital image comprising a plurality of pixels, and wherein a position of each pixel in the plurality of pixels of the spherical panorama digital image reflects an incoming light direction corresponding to light rays portrayed in the input digital image.