$2.$ Image personalization with Textual Inversion

In this task your goal is to generate images containing specific object, character, pet or a person at your preference.

Textual Inversion is a training technique for personalizing image generation models with just a few example images of what you want it to learn. This technique works by learning and updating the text embeddings $($the new embeddings are tied to a special word you must use in the prompt$)$ to match the example images you provide.

Data preparation

Prepare your dataset with $5-10($or more$)$ images with the object or concept of interest. You may adopt TextualInversionDataset from diffusers for training.

Training

Train the model and visualize the results on $5-10$ prompts.

Ablation

Experiment with the number of learnable embeddings. How does it affect quality? Try $3-5$ options.

Ablate the dataset size. Does a single image suffice for a good quality. How many samples are needed? Try $3-5$ different options. Visualize the results for some $5$ prompts and each of the option.