Pokémon Generation Diffusion Model

A Pokémon generator trained using the DDPM forward process and performing the reverse process with both DDPM and stochastic DDIM sampling.

---

Dataset

The dataset used for training is located at data/archive.zip. It was retrieved from Kaggle's Pokémon Sprites Dataset.

Key Details:

• Size: 10,437 images
• Resolution: 96x96
• Classes: 898 Pokémon across various games

---

Model

The model is a U-Net trained with the DDPM forward process for 50 epochs.

Artifacts:

• Trained Model: The trained state_dict is available at model/pokemon.pth.
• Code: The full implementation is in Diffusion_updated.ipynb.

---

Sample Outputs

The generated samples are resized to 64x64 for display purposes.

Below are examples of outputs from the model. Empirically, DDIM sampling tends to produce better results compared to DDPM sampling:

---

Generate Your Own Pokémon!

Follow these steps to generate your own Pokémon:

1. Clone this repository to your local environment
2. Create a Python virtual environment and install torch and matplotlib in it
3. Open the folder by running cd pokemon_generator_diffusion
4. To use the DDPM sampler, run make generate_ddpm To use the DDIM sampler, run make generate_ddim

• Time: The generation process typically takes 1 to 2 minutes, depending on your device.
• Occasional Issues: The model may sometimes produce completely white images due to occasional performance issues. If this happens, simply run the generation process again.

Potential Next Step

Current Limitation:

The model was trained with DDPM, so using DDIM with intervals larger than 1 can result in unstable or noisy outputs since the model didn't learn how to handle larger timesteps.

Future Directions:

• Fine-tuning the model with DDIM: Adapt the model to handle larger intervals for faster sampling.
• Experimenting with stochastic DDIM: Improve stability and performance by tuning the noise schedule (η) or experimenting with hybrid sampling methods.