Ronenk94/T5_matryoshka_sae_top_300

🧠 Top-K 300 Sparse Autoencoder (SAE) — SAEdit

Repo: Ronenk94/T5_matryoshka_sae
Model Type: Sparse Autoencoder over T5 Embeddings
Paper: SAEdit: Token-level control for continuous image editing via Sparse AutoEncoder
License: CC BY 4.0

---

📌 Model Overview

This repository contains the Top-K 300 Sparse Autoencoder (SAE) used in the SAEdit framework.
It is trained on T5 text embeddings and designed to produce sparse latent representations that enable token-level semantic control in image editing pipelines.

Property	Details
Architecture	GlobalBatchTopKMatryoshkaSAE
Latent sparsity	Top-K = 300 activations
Backbone embeddings	Frozen T5 encoder
Task	Semantic factorization + reconstruction
Use case	Editing directions for diffusion-based image manipulation

---

📥 How to Load

import torch
from src.models.sparse_autoencoders.matryoshka_sae import GlobalBatchTopKMatryoshkaSAE

# Option A — using a from_pretrained method (if implemented)
model = GlobalBatchTopKMatryoshkaSAE.from_pretrained(
    "Ronenk94/T5_matryoshka_sae",
    device="cuda"
)

📄 Citation

If you use this model for your research, please cite the following work:

@misc{kamenetsky2025saedittokenlevelcontrolcontinuous,
      title={SAEdit: Token-level control for continuous image editing via Sparse AutoEncoder}, 
      author={Ronen Kamenetsky and Sara Dorfman and Daniel Garibi and Roni Paiss and Or Patashnik and Daniel Cohen-Or},
      year={2025},
      eprint={2510.05081},
      archivePrefix={arXiv},
      primaryClass={cs.GR},
      url={https://arxiv.org/abs/2510.05081}, 
}

⚠️ Limitations & Risks

Model was trained on T5 encoder of the Flux-Dev variant. Other models might use different checkpoint

Lisence

This model is released under Creative Commons Attribution 4.0 (CC BY 4.0), consistent with the associated SAEdit paper.