[MM'22 Oral] AI Illustrator: Translating Raw Descriptions into Images by Prompt-based Cross-Modal Generation
This reposity is the official implementation of AI Illustrator: Translating Raw Descriptions into Images by Prompt-based Cross-Modal Generation.
The proposed pipeline is shown below.
Currently, we support pretrained models on 3 domains: Human face, Cat and Church. The download urls are:
| Human face | Cat | Church | |
|---|---|---|---|
| Baidu Disk | https://pan.baidu.com/s/1wDVj_YGYQoQlRDk5tTtL8A, extracting code: huzx |
https://pan.baidu.com/s/1UIhCBL2Cl9CenjjCNsbHJw, extracting code: 7ul6 |
https://pan.baidu.com/s/1WTrWgrjs9FD8o4ZyoQalxg, extracting code: hffh |
| Google Drive | TBD | TBD | TBD |
| One Drive | https://1drv.ms/u/s!Aq9epwaFQGaOgQpUdtd81YWh_TVe?e=5vuPjD | https://1drv.ms/u/s!Aq9epwaFQGaOgQwjbTUXjYgw-g_z?e=lTg9lX | https://1drv.ms/u/s!Aq9epwaFQGaOgQ1foRYeby5jHR96?e=g4yq8G |
The default path of pretrained models is ./pretrained_projectors.
After downloading the pretrained models, you can simply generate images by command
python single_generate.py --kind <domain> --projector_path <path/to/the/pretrained_projector> --save_path <path/to/the/save_dir> --strength 1.75 --prompt_path <path/to/the/text_prompt>
One example is
python single_generate.py --kind 'human' --projector_path './pretrained_projector/c2s_human.pth' --save_path './outputs' --strength 1.75 --prompt_path './prompts/ffhq_text_prompt.pth'
The values of argument "kind", "projector_path" and "prompt_path" should match. By default, "kind" should be one of "human", "cat" and "church".
100 raw descriptions can be found in ./benchmark/description_benchmark_100.txt
Corresponding generation results of our method can be download at
| Baidu Disk | Google Drive | One Drive |
|---|---|---|
| https://pan.baidu.com/s/1yNrQs5MxbUkKDDUw0rv4Wg, extracting code: 0y0s |
TBD | https://1drv.ms/u/s!Aq9epwaFQGaOgQ4vXrB4rZqZkHOE?e=BHEaMD |