System and method for generating photorealistic synthetic images based on semantic information
Embodiments described herein provide a system for generating semantically accurate synthetic images. During operation, the system generates a first synthetic image using a first artificial intelligence (AI) model and presents the first synthetic image in a user interface. The user interface allows a user to identify image units of the first synthetic image that are semantically irregular. The system then obtains semantic information for the semantically irregular image units from the user via the user interface and generates a second synthetic image using a second AI model based on the semantic information. The second synthetic image can be an improved image compared to the first synthetic image.
1. A method for generating semantically accurate synthetic images, comprising:
generating a first synthetic image using a first artificial intelligence (AI) model;
presenting the first synthetic image in a user interface, wherein the user interface is configured to receive information indicating image units of the first synthetic image that are semantically irregular;
obtaining semantic information for the semantically irregular image units from a user via the user interface;
incorporating the semantic information into a loss function of a second AI model; and
generating a second synthetic image using the second AI model based on the loss function, wherein the second synthetic image is generated when the loss function is within a threshold and includes one or more improved image units compared to the first synthetic image.
2. The method of claim 1 , further comprising:
obtaining a selection of a highlighting tool for the user interface, wherein a region marked by the highlighting tool is selectable via the user interface, and wherein the highlighting tool corresponds to one of: a grid-based selector, a polygon-based selector, and a freehand selector; and
in response to the highlighting tool being the grid-based selector, setting the highlighting tool at an obtained granularity.
3. The method of claim 2 , further comprising:
obtaining a selection of a region selected by the highlighting tool in the user interface; and
obtaining a weight allocated for the region.
4. The method of claim 1 , wherein the first and second AI models are generative adversarial networks (GANs).
5. The method of claim 1 , wherein the second AI model includes a discriminator that outputs a spatial probability map, wherein a respective element of the spatial probability map indicates the probability of an image unit of the second synthetic image being realistic.
6. The method of claim 5 , wherein the second AI model includes a generator that outputs the second synthetic image in such a way that corresponding image units, which include the one or more improved image units, are different than the semantically irregular image units of the first synthetic image.
7. The method of claim 6 , further comprising generating a spatial weight mask based on the obtained semantic information, wherein the spatial weight mask comprises weights allocated to a respective image unit of the first synthetic image, and wherein the generator and the discriminator determine the semantically irregular image units based on the spatial weight mask.
8. The method of claim 1 , wherein an image unit corresponds to one or more of: a pixel, a pixel block, and a group of pixels of an irregular shape.
9. The method of claim 1 , wherein generating the first and the second synthetic images includes mapping a noise vector into a synthetic image sample.
10. The method of claim 1 , further comprising:
generating a third synthetic image using the first AI model;
automatically obtaining semantic information for semantically irregular image units of the third synthetic image based on an irregularity detection technique; and
generating a fourth synthetic image using the second AI model based on the automatically obtained semantic information.
11. A non-transitory computer-readable storage medium storing instructions that when executed by a computer cause the computer to perform a method for generating semantically accurate synthetic images, the method comprising:
generating a first synthetic image using a first artificial intelligence (AI) model;
presenting the first synthetic image in a user interface, wherein the user interface is configured to receive information indicating image units of the first synthetic image that are semantically irregular;
obtaining semantic information for the semantically irregular image units from a user via the user interface;
incorporating the semantic information into a loss function of a second AI model; and
generating a second synthetic image using the second AI model based on the loss function, wherein the second synthetic image is generated when the loss function is within a threshold and includes one or more improved image units compared to the first synthetic image.
12. The non-transiory computer-readable storage medium of claim 11 , wherein the method further comprises:
obtaining a selection of a highlighting tool for the user interface, wherein a region marked by the highlighting tool is selectable via the user interface, and wherein the highlighting tool corresponds to one of: a grid-based selector, a polygon-based selector, and a freehand selector; and
in response to the highlighting tool being the grid-based selector, setting the highlighting tool at an obtained granularity.
13. The non-transitory computer-readable storage medium of claim 12 , wherein the method further comprises:
obtaining a selection of a region selected by the highlighting tool in the user interface; and
obtaining a weight allocated for the region.
14. The non-transitory computer-readable storage medium of claim 11 , wherein the first and second AI models are generative adversarial networks (GANs).
15. The non-transitory computer-readable storage medium of claim 11 , wherein the second AI model includes a discriminator that outputs a spatial probability map, wherein a respective element of the spatial probability map indicates the probability of an image unit of the second synthetic image being realistic.
16. The non-transitory computer-readable storage medium of claim 15 , wherein the second AI model includes a generator that outputs the second synthetic image in such a way that corresponding image units are different than the semantically irregular image units of the first synthetic image.
17. The non-transitory computer-readable storage medium of claim 16 , wherein the method further comprises generating a spatial weight mask based on the obtained semantic information, wherein the spatial weight mask comprises weights allocated to a respective image unit of the first synthetic image, and wherein the generator and the discriminator determine the semantically irregular image units based on the spatial weight mask.
18. The non-transitory computer-readable storage medium of claim 11 , wherein an image unit corresponds to one or more of: a pixel, a pixel block, and a group of pixels of an irregular shape.
19. The non-transitory computer-readable storage medium of claim 11 , wherein generating the first and the second synthetic images includes mapping a noise vector into a synthetic image sample.
20. The non-transitory computer-readable storage medium of claim 11 , wherein the method further comprises:
generating a third synthetic image using the first AI model;
automatically obtaining semantic information for semantically irregular image units of the third synthetic image based on an irregularity detection technique; and
generating a fourth synthetic image using the second AI model based on the automatically obtained semantic information.