Github Repo The detailed work description and code can be found in https://github.com/foxintohumanbeing/DDA4210_Group_project.
The creation of high-quality image content from text descriptions is a challenging yet highly desirable task in the field of artificial intelligence. We focus on the Simpsons, a popular animated series. Based on pretrained SOTA model, we investigate in obtaining high-quality dataset and efficient fine-tuning methods. We explore the options of manually creating the dataset and using different fine-tuning techniques such as simple baseline, LoRA, and Dreambooth. Our approach involves combining the advantages of each option to achieve better results.
We propose dataset collection method and fine-tuning model(Simspon Artistic Memory). Moreover, to better illustrating our results, we create two APPs, one for generating images and one for annotating the images (you can find them in github link provided). By improving data collection and fine-tuning techniques on Simpsons, we hope to push the boundaries of what is achievable in the text-to-image synthesis domain and inspire further research in this area.
Prompts Format "Asim. a [closeup?] of a [emotional expression] [race] [X year old] [man / woman / etc.], with [hair and makeup style], wearing [clothing style] while [doing] near [nearby objects],[outside / inside] with [objects / color ] in the background,in [time period]."
Contact
For any questions, please contact me at 120090214@link.cuhk.edu.cn