Training Diffusion Models with Reinforcement Learning
Paper
• 2305.13301 • Published
• 5
DDPO BRISQUE Checkpoints
This repository contains checkpoints for Stable Diffusion v1.5 fine-tuned using Denoising Diffusion Policy Optimization (DDPO).
Training Details
- Base Model: Stable Diffusion v1.5
- Reward Function: BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator)
- Method: LoRA-based DDPO
- Training Platform: MIT ORCD (Engaging)
Checkpoints
We provide checkpoints every 10 epochs plus the final epoch (98). BRISQUE is a no-reference image quality metric where lower values typically indicate higher spatial quality/less distortion.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support