Generating large datasets of style-specific and content-specific images using generative machine-learning models to match a small set of sample images
The present disclosure relates to utilizing a style-matching image generation system to generate large datasets of style-matching images having matching styles and content to an initial small sample set of input images. For example, the style-matching image generation system utilizes a selection of style-mixed stored images with a generative machine-learning model to produce large datasets of synthesized images. Further, the style-matching image generation system utilizes the generative machine-learning model to conditionally sample synthesized images that accurately match the style, content, characteristics, and patterns of the initial small sample set and that also provide added variety and diversity to the large image dataset.
1. A computer-implemented method, comprising:
comparing an initial set of input images to multiple large sets of stored image datasets to determine a style distribution between the initial set of input images and the multiple large sets of stored image datasets;
determining an expanded set of input images by selecting images from one or more stored image sets of the multiple large sets of stored image datasets in accordance with the style distribution; and
generating a set of style-matching images utilizing a generative machine-learning model by:
generating a set of synthesized images based on the expanded set of input images; and
conditionally sampling the set of style-matching images from the set of synthesized images based on the initial set of input images.
2. The computer-implemented method of claim 1 , wherein the expanded set of input images comprises the initial set of input images and the images selected from the one or more stored image sets.
3. The computer-implemented method of claim 1 , wherein comparing the initial set of input images to the multiple large sets of stored image datasets comprises:
generating a first set of image embeddings based on the initial set of input images;
generating a second set of image embeddings based on a first stored image set of the multiple large sets of stored image datasets; and
generating a third set of image embeddings based on a second stored image set of the multiple large sets of stored image datasets.
4. The computer-implemented method of claim 3 , wherein comparing the initial set of input images to the multiple large sets of stored image datasets further comprises:
determining a first embedding distance between the first set of image embeddings and the second set of image embeddings;
determining a second embedding distance between the first set of image embeddings and the third set of image embeddings; and
determining, based on comparing the first embedding distance to the second embedding distance, the style distribution comprising a first style distribution value for the first stored image set and a second style distribution value for the second stored image set.
5. The computer-implemented method of claim 1 , wherein determining the expanded set of input images by selecting images from the one or more stored image sets in accordance with the style distribution comprises sampling stored image datasets from the one or more stored image sets across a probabilistic distribution in accordance with a style distribution value determined for each of the one or more stored image sets.
6. The computer-implemented method of claim 1 , wherein determining the expanded set of input images by selecting images from the one or more stored image sets in accordance with the style distribution comprises sampling stored image datasets in a given stored image set based on minimizing an image distance between images in the given stored image set and the initial set of input images.
7. The computer-implemented method of claim 1 , wherein conditionally sampling the set of style-matching images from the set of synthesized images comprises utilizing the generative machine-learning model to generate the set of style-matching images based on sampling images that correspond to the initial set of input images from an embedding space corresponding to the expanded set of input images.
8. The computer-implemented method of claim 1 , wherein generating the set of synthesized images based on the expanded set of input images comprises training the generative machine-learning model based on the expanded set of input images to generate a trained generative machine-learning model.
9. The computer-implemented method of claim 8 , wherein conditionally sampling the set of style-matching images from the set of synthesized images comprises generating the set of style-matching images with the trained generative machine-learning model utilizing the initial set of input images as a conditional input.
10. The computer-implemented method of claim 1 , further comprising utilizing the set of style-matching images to train an image-based machine-learning model.
11. A system comprising:
at least one processor; and
a computer memory comprising instructions that, when executed by the at least one processor, cause the system to carry out operations comprising:
comparing an initial set of input images to multiple sets of stored images to determine a style distribution between the initial set of input images and the multiple sets of stored images;
determining an expanded set of input images by selecting images from one or more stored image sets of the multiple sets of stored images in accordance with the style distribution; and
generating a set of style-matching images utilizing a generative machine-learning model by:
generating a set of synthesized images based on the expanded set of input images; and
conditionally sampling the set of style-matching images from the set of synthesized images based on the initial set of input images.
12. The system of claim 11 , further comprising instructions that, when executed by the at least one processor, cause the system to carry out operations comprising:
receiving the initial set of input images from a client device; and
providing the set of style-matching images to the client device in response to receiving the initial set of input images.
13. The system of claim 11 , further comprising instructions that, when executed by the at least one processor, cause the system to carry out operations comprising:
determining a first embedding distance between a first set of image embeddings corresponding to the initial set of input images and a second set of image embeddings corresponding to a first stored image set of the multiple sets of stored images;
determining a second embedding distance between the first set of image embeddings and a third set of image embeddings corresponding to a second stored image set of the multiple sets of stored images; and
determining a first style distribution value for the first stored image set and a second style distribution value for the second stored image set based on comparing the first embedding distance to the second embedding distance.
14. The system of claim 13 , wherein the first style distribution value for the first stored image set is larger than the second style distribution value for the second stored image set based on the first stored image set having a closer image style to a target image style of the initial set of input images.
15. The system of claim 14 , wherein the first style distribution value and the second style distribution value add up to one.
16. A computer-implemented method comprising:
comparing an initial set of input images to multiple sets of stored images to determine a set of style distribution values between the initial set of input images and the multiple sets of stored images;
determining an expanded set of input images by sampling images from a plurality of the multiple sets of stored images in accordance with the set of style distribution values;
generating a set of synthesized images utilizing a first generative machine-learning model based on the expanded set of input images; and
generating a set of style-matching images by conditionally sampling the set of style-matching images from the set of synthesized images utilizing a second generative machine-learning model.
17. The computer-implemented method of claim 16 , wherein the set of style-matching images closely aligns image style and an image content to the initial set of input images in style and content.
18. The computer-implemented method of claim 16 , wherein the first generative machine-learning model differs from the second generative machine-learning model.
19. The computer-implemented method of claim 16 , wherein the first generative machine-learning model and the second generative machine-learning model are a same generative machine-learning model.
20. The computer-implemented method of claim 16 , wherein generating the set of style-matching images comprises utilizing the second generative machine-learning model to generate the set of style-matching images based on sampling generated images within an embedding space having image embeddings generated by the second generative machine-learning model that have a shortest embedding distance from embeddings of the initial set of input images.