Self-Play Fine-Tuning of Diffusion Models for Text-to-Image Generation (https://huggingface.co/papers/2402.10210)

SPIN-Diffusion-iter1

This model is a self-play fine-tuned diffusion model at iteration 1 from runwayml/stable-diffusion-v1-5 using synthetic data based on the winner images of the yuvalkirstain/pickapic_v2 dataset. We have also made a Gradio Demo at UCLA-AGI/SPIN-Diffusion-demo-v1.

Model Details

Model Description

• Model type: An diffusion model with unet fine-tuned, based on the strucure of stable diffusion 1.5
• Language(s) (NLP): Primarily English
• License: Apache-2.0
• Finetuned from model: runwayml/stable-diffusion-v1-5

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2.0e-05
• train_batch_size: 8
• distributed_type: multi-GPU
• num_devices: 8
• train_gradient_accumulation_steps: 32
• total_train_batch_size: 2048
• optimizer: AdamW
• lr_scheduler: "linear"
• lr_warmup_steps: 200
• num_training_steps: 500

Citation

@misc{yuan2024self,
      title={Self-Play Fine-Tuning of Diffusion Models for Text-to-Image Generation}, 
      author={Yuan, Huizhuo and Chen, Zixiang and Ji, Kaixuan and Gu, Quanquan},
      year={2024},
      eprint={2402.10210},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}