README
The Universal Normal Embedding (UNE)
Official code for The Universal Normal Embedding (arXiv:2603.21786), studying how unguided diffusion represents images.
What this repository is about
We advance the Universal Normal Embedding (UNE) hypothesis: that there is a shared, Gaussian latent representation underlying both generative models and image encoders—two lines of work that are usually treated separately, but may be noisy linear views of the same underlying space.
On the empirical side, we show that the initial noise used in unguided diffusion is not semantically inert: it carries structured, rankable information about the image that will be generated. For example, a linear classifier fit on the starting noise can predict high-level attributes of the output—such as whether the generated face will read as male or female—without looking at the final pixels. That pattern extends to other attributes and settings described in the paper.
Subsequently, we use those linear classifiers to edit generated images by manipulating the initial noise, and apply geometric techniques to identify editing directions orthogonal to undesired attribute changes, thereby removing spurious attribute changes during editing.
Links
| Resource | Link |
|---|---|
| Paper | Paper on arXiv |
| Project website | UNE project page |
| Dataset | NoiseZoo on Hugging Face |
Repository layout
utils/— Training linear probes (LDA, logistic regression, etc.), data helpers, diffusion pipeline wrappers, and utilities for linear editing in latent / noise space (including orthogonalization and optional gaussianization for interpretable edit strength).examples/— Notebook walkthroughs (seeexamples/README.md).
Example notebooks
| Notebook | What it demonstrates |
|---|---|
examples/classification_demo.ipynb | Linear classification on latent representations (e.g. CelebA attributes). |
examples/editing_demo.ipynb | Linear editing by moving latents along directions learned from linear probes (e.g. attribute directions). |
GPU is recommended for decoding and diffusion-related steps.
Setup
Python 3 with PyTorch is assumed. From the repository root:
pip install -r requirements.txt
The requirements.txt pins CUDA-oriented PyTorch builds (see --extra-index-url inside the file); PyTorch can still run on CPU when no GPU is available.
Running the notebooks
From the project root:
jupyter notebook examples/classification_demo.ipynb
The notebooks add the parent directory to sys.path so utils imports resolve when run from examples/.
Citation
If you use this code or the associated dataset, please cite the paper.
License
See LICENSE.
Contributors
3 commits
Similar repos
Languages
Python
100.0%
README
The Universal Normal Embedding (UNE)
Official code for The Universal Normal Embedding (arXiv:2603.21786), studying how unguided diffusion represents images.
What this repository is about
We advance the Universal Normal Embedding (UNE) hypothesis: that there is a shared, Gaussian latent representation underlying both generative models and image encoders—two lines of work that are usually treated separately, but may be noisy linear views of the same underlying space.
On the empirical side, we show that the initial noise used in unguided diffusion is not semantically inert: it carries structured, rankable information about the image that will be generated. For example, a linear classifier fit on the starting noise can predict high-level attributes of the output—such as whether the generated face will read as male or female—without looking at the final pixels. That pattern extends to other attributes and settings described in the paper.
Subsequently, we use those linear classifiers to edit generated images by manipulating the initial noise, and apply geometric techniques to identify editing directions orthogonal to undesired attribute changes, thereby removing spurious attribute changes during editing.
Links
| Resource | Link |
|---|---|
| Paper | Paper on arXiv |
| Project website | UNE project page |
| Dataset | NoiseZoo on Hugging Face |
Repository layout
utils/— Training linear probes (LDA, logistic regression, etc.), data helpers, diffusion pipeline wrappers, and utilities for linear editing in latent / noise space (including orthogonalization and optional gaussianization for interpretable edit strength).examples/— Notebook walkthroughs (seeexamples/README.md).
Example notebooks
| Notebook | What it demonstrates |
|---|---|
examples/classification_demo.ipynb | Linear classification on latent representations (e.g. CelebA attributes). |
examples/editing_demo.ipynb | Linear editing by moving latents along directions learned from linear probes (e.g. attribute directions). |
GPU is recommended for decoding and diffusion-related steps.
Setup
Python 3 with PyTorch is assumed. From the repository root:
pip install -r requirements.txt
The requirements.txt pins CUDA-oriented PyTorch builds (see --extra-index-url inside the file); PyTorch can still run on CPU when no GPU is available.
Running the notebooks
From the project root:
jupyter notebook examples/classification_demo.ipynb
The notebooks add the parent directory to sys.path so utils imports resolve when run from examples/.
Citation
If you use this code or the associated dataset, please cite the paper.
License
See LICENSE.
Similar repos
Contributors
3 commits
Languages
Python
100.0%