zhaoyl18/SEIKO

SEIKO is a novel reinforcement learning method to efficiently fine-tune diffusion models in an online setting. Our methods outperform all baselines (PPO, classifier-based guidance, direct reward backpropagation) for fine-tuning Stable Diffusion.

23
/ 100
Experimental

This project helps researchers and developers fine-tune diffusion models to generate data (like images, biological sequences, or molecules) that meet specific, high-value criteria. You provide a pre-trained diffusion model and a way to evaluate the 'goodness' of generated samples, and it outputs a refined model capable of consistently producing high-reward results. This is for machine learning practitioners and scientists working on generative AI for specialized applications.

No commits in the last 6 months.

Use this if you need to efficiently train a diffusion model to generate items with a specific, measurable property, such as aesthetically pleasing images or molecules with high bioactivity, especially when evaluating these properties can be costly or time-consuming.

Not ideal if your goal is general-purpose data generation without a specific, optimizable 'reward' metric, or if you do not have the infrastructure (like GPUs) to train large generative models.

generative-AI materials-discovery drug-design bioinformatics computational-chemistry
Stale 6m No Package No Dependents
Maintenance 0 / 25
Adoption 7 / 25
Maturity 16 / 25
Community 0 / 25

How are scores calculated?

Stars

30

Forks

Language

Python

License

MIT

Last pushed

Jul 18, 2024

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/diffusion/zhaoyl18/SEIKO"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.