MyDreamer: Personalized Text-to-Vecor Generation with DiffusionModel

디퓨전을 활용한 개인화된 텍스트 기반 벡터 그래픽 생성 모델

Environment Setup

Clone the repository:

git clone https://github.com/soyeong-kwon/MyDreamer.git
sh script/install.sh

Pretrained Weights

Download the personalized model weights from DreamMatcher. These weights are pretrained DreamBooth weights on the ViCo dataset.
You can find the pre-trained weights Link.

After downloading, store the weights in the following directory:

dreambooth/concept_models/

Running the Model

To run the model, execute:

sh script/run.sh

Customization Options

You can modify the following parameters to customize your settings:

• num_batch: Batch size for generating graphics.
• prompt_index_from / prompt_index_to: Specify the range of prompt indices you want to generate.
• mode: Set to normal or challenging.
• concept_id: Define which object you want to generate.
• total_step: Total number of steps for optimization.
• flag_prev: Set to true to enable DVSD.
• flag_igs: Set to true to enable DGS.
• gs_type: Define the DGS scheduling scheme.
• bg_lambda: Lambda value for background optimization.
• mprec: Use no for float32 precision or fp16 for float16 precision.

Evaluation

To evaluate the results, run:

sh script/eval.sh

Before you run

Please update the following paths in the script:

• real_folder: Path to the folder containing real pixel images. (for FID)
• sample_folder: Path to the folder containing sampled images.
• sample_dm_folder: Path to the folder containing sampled images.

Note: The sample_folder and sample_dm_folder contain the same sampled images but differ in folder structure. You can use utils/move_samples.py to create the sample_folder.