aim-uofa

aim-uofa / MARBLE

Public

Multi-Aspect Reward Balance for Diffusion RL

19
0
100% credibility
Found May 08, 2026 at 19 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
AI Summary

MARBLE is a research project offering pre-trained models and tools to fine-tune AI image generators for balancing multiple quality rewards simultaneously.

How It Works

1
🔍 Discover MARBLE

You hear about this exciting project that helps AI create images great at many things like looks, readability, and more—all at once.

2
📖 Explore the project

You check out the page and paper to see how it balances different goals without one overpowering the others.

3
⭐ Grab ready-made models

You download the pre-trained image creators that let you start generating right away.

4
🎨 Create your images

You describe what you want, and it makes pictures that shine in multiple ways without trade-offs.

5
📊 Check the results

You see how your images score high on beauty, clarity, and other qualities together.

✨ Celebrate balanced masterpieces

Now you have stunning images that excel everywhere, ready for your projects or fun.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 19 to 19 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is MARBLE?

MARBLE fine-tunes diffusion models with multiple rewards in RL, balancing them end-to-end in gradient space to preserve specialist signals like OCR or aesthetics. It solves the dilution from scalar averaging or sequential training, letting one model handle diverse objectives simultaneously—think download blue marble github for diffusion balance, not github marble blast games. Built on DiffusionNFT with Hugging Face checkpoints for inference now, training code drops later.

Why is it gaining traction?

It trains jointly across rewards at near single-reward cost, boosting composite scores +1.12 over baselines while specialist metrics hold up—no more OCR collapse. Developers hook on the project page's curves showing stable multi-aspect gains, unlike marble run github toys or kde marble github maps. Ties into marbles on stream workflows for balanced diffusion RL.

Who should use this?

ML engineers at research labs like Zhejiang University optimizing text-to-image diffusion for multi-reward RLHF, e.g., visual appeal plus text fidelity. Teams building marble mountains of quality in gen AI, avoiding marble rush tradeoffs. Not for marble cms github sites or marble deutsch localization.

Verdict

Early days at 19 stars and 1.0% credibility—only README, inference checkpoints, no training code yet, docs lean on arXiv and project page. Star for diffusion RL multi-reward needs; hold off if maturity matters.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.