valeoai

valeoai / REPA-G

Public

Official Repositary of "Test-Time Conditioning with Representation-Aligned Visual Features"

17
0
100% credibility
Found Feb 06, 2026 at 16 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Jupyter Notebook
AI Summary

REPA-G enables test-time image generation conditioned on visual features from pretrained encoders like DINOv2 using SiT diffusion models, with a Streamlit demo and evaluation tools.

How It Works

1
🔍 Discover REPA-G

You stumble upon this exciting image generator that creates pictures by blending ideas from your own photos, no text needed.

2
📦 Get ready

Download a simple list of helpers and add them to your computer with one easy command.

3
⬇️ Grab the magic

Fetch the pre-trained brains that make the generator smart with a quick download script.

4
🚀 Launch the playground

Start the friendly app in your web browser to play with image blending right away.

5
🖼️ Mix and create

Pick your photos, choose what to blend like textures or shapes, adjust strength sliders, and watch new images appear instantly.

✨ Your custom art

Enjoy beautiful, personalized images that perfectly capture the essence of your reference photos, ready to save and share.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 16 to 17 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 REPA-G?

REPA-G is an official GitHub repository implementing test-time conditioning for diffusion models using representation-aligned visual features. It lets you guide image generation at inference by steering denoising towards semantic features from pre-trained extractors like DINOv2 or SiT internals, enabling precise control from patch-level textures to global concepts or multi-object compositions. Built in Python with PyTorch, it runs via scripts or a Streamlit demo, producing high-quality ImageNet/COCO samples without retraining—ideal as a text-prompt alternative.

Why is it gaining traction?

It stands out by working purely at test time, avoiding prompt ambiguity for faithful visual fidelity, with scripts for eval on ImageNet/COCO including FID, CLIP scores, and custom discriminability metrics. Developers appreciate the drop-in integration with pretrained SiT/REPA-E models, batched generation, and extras like PCA-masked potentials for focused editing. The toy Jupyter example and download scripts make prototyping fast.

Who should use this?

AI researchers tuning controllable diffusion for semantic editing, like matching specific ImageNet styles or composing COCO scenes. Computer vision engineers needing precise visual guidance over text, such as in AR/VR asset gen or data augmentation pipelines. Teams extending REPA baselines for custom backbones.

Verdict

Grab it if you're in diffusion research—solid paper-backed features and evals despite 17 stars and 1.0% credibility score signaling early-stage maturity. Docs are README-focused with setup scripts; expect tweaks for production, but the Streamlit demo hooks quick experiments.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.