README
NoiseZoo: CelebA Latents & Diffusion Inversion Noises
NoiseZoo is a unified collection of latent representations and diffusion inversion noises for the CelebA dataset. It provides a "Rosetta Stone" for vision research, mapping the same image set across 10 different state-of-the-art encoder and generative architectures.
Quick Links
Dataset Composition
The dataset consists of 10 .npy files. Each file is a 2D NumPy matrix of shape (162770, D), where each row corresponds to the standard CelebA training image index.
Model Specifications
All files are located in the data/ directory. All representations have been flattened into 1D vectors per image.
| Model Name | File Name | Dimension ($D$) | Type |
|---|---|---|---|
| Stable Diffusion 1.5 | noises_sd15.npy | 16384 (4x64x64) | DDIM Inversion |
| Stable Diffusion 2.1 | noises_sd21.npy | 16384 (4x64x64) | DDIM Inversion |
| SDXL | noises_sdxl.npy | 65536 (4x128x128) | DDIM Inversion |
| LCM v7 | noises_lcm.npy | 16384 | DDIM Inversion |
| DINO | noises_dino.npy | 1024 | Encoder Latent |
| CLIP B/16 | noises_clipB16.npy | 512 | Encoder Latent |
| CLIP L/14 | noises_clipL14.npy | 768 | Encoder Latent |
| ConvNeXT | noises_convnext.npy | 768 | Encoder Latent |
| MAE | noises_mae.npy | 768 | Encoder Latent |
| Swin | noises_swin.npy | 768 | Encoder Latent |
Inversion Parameters
| Parameter | SD 1.5, SD 2.1, LCM v7 | SDXL |
|---|---|---|
| DDIM steps | 150 | 300 |
| Resolution | 512² | 1024² |
| Latent (C×H×W) | 4×64×64 | 4×128×128 |
| Scheduler | DDIM (from_config) | DDIM (from_config) |
| Prompt | "" | "" (inversion without CFG) |
Usage
Loading via Hugging Face
The following snippet demonstrates how to load a specific noise file into a NumPy array:
import numpy as np
from huggingface_hub import hf_hub_download
# Downloads the specific file and returns the local path to it
file_path = hf_hub_download(
repo_id="chentasker/NoiseZoo",
filename="data/noises_clipB16.npy",
repo_type="dataset"
)
# Now load it normally with numpy
latents = np.load(file_path)
print(f"Success! Matrix shape: {latents.shape}")
Contributors
9 commits
Similar repos
Linked in READMEs
README
NoiseZoo: CelebA Latents & Diffusion Inversion Noises
NoiseZoo is a unified collection of latent representations and diffusion inversion noises for the CelebA dataset. It provides a "Rosetta Stone" for vision research, mapping the same image set across 10 different state-of-the-art encoder and generative architectures.
Quick Links
Dataset Composition
The dataset consists of 10 .npy files. Each file is a 2D NumPy matrix of shape (162770, D), where each row corresponds to the standard CelebA training image index.
Model Specifications
All files are located in the data/ directory. All representations have been flattened into 1D vectors per image.
| Model Name | File Name | Dimension ($D$) | Type |
|---|---|---|---|
| Stable Diffusion 1.5 | noises_sd15.npy | 16384 (4x64x64) | DDIM Inversion |
| Stable Diffusion 2.1 | noises_sd21.npy | 16384 (4x64x64) | DDIM Inversion |
| SDXL | noises_sdxl.npy | 65536 (4x128x128) | DDIM Inversion |
| LCM v7 | noises_lcm.npy | 16384 | DDIM Inversion |
| DINO | noises_dino.npy | 1024 | Encoder Latent |
| CLIP B/16 | noises_clipB16.npy | 512 | Encoder Latent |
| CLIP L/14 | noises_clipL14.npy | 768 | Encoder Latent |
| ConvNeXT | noises_convnext.npy | 768 | Encoder Latent |
| MAE | noises_mae.npy | 768 | Encoder Latent |
| Swin | noises_swin.npy | 768 | Encoder Latent |
Inversion Parameters
| Parameter | SD 1.5, SD 2.1, LCM v7 | SDXL |
|---|---|---|
| DDIM steps | 150 | 300 |
| Resolution | 512² | 1024² |
| Latent (C×H×W) | 4×64×64 | 4×128×128 |
| Scheduler | DDIM (from_config) | DDIM (from_config) |
| Prompt | "" | "" (inversion without CFG) |
Usage
Loading via Hugging Face
The following snippet demonstrates how to load a specific noise file into a NumPy array:
import numpy as np
from huggingface_hub import hf_hub_download
# Downloads the specific file and returns the local path to it
file_path = hf_hub_download(
repo_id="chentasker/NoiseZoo",
filename="data/noises_clipB16.npy",
repo_type="dataset"
)
# Now load it normally with numpy
latents = np.load(file_path)
print(f"Success! Matrix shape: {latents.shape}")
Similar repos
Linked in READMEs
Contributors
9 commits