This repo compares images generated by awesome generative models on the DrawBench provided in Imagen. Currently, it includes images generated with some OpenSource image generative models (stable diffusion 1-4, 1-5, 2-1). One can browse the generated images and get a rough idea about the capability of these models.
- The repo includes results generated by 3 stable diffusion models (1-4, 1-5, 2-1) with 200 prompts from DrawBench, organized by prompt categories.
- The provided scripts can be used to generate images with additional prompts and private models and provide organized results (the README files in each sub-directory display the results).
- Currently, this repo only includes 3 OpenSource stable diffusion models.
- Currently, this repo only includes images generated with one random seed (
0
), which makes the comparison not comprehensive.
- Create a conda environment:
conda env create -f bench.yaml
. - Activate the conda environment:
conda activate bench
. - Run the notebook
GenImages.ipynb
. It takes ~30 minutes to generate a single image for 200 prompts with 3 different diffusion models on a single RTX3090.
Each sub-directory includes the generation results of a single prompt category. Some examples are provided below as a preview.