|
|
--- |
|
|
title: DiffViews |
|
|
emoji: 🔬 |
|
|
colorFrom: purple |
|
|
colorTo: blue |
|
|
sdk: gradio |
|
|
sdk_version: 6.3.0 |
|
|
app_file: app.py |
|
|
python_version: "3.10" |
|
|
pinned: false |
|
|
license: mit |
|
|
--- |
|
|
|
|
|
# DiffViews - Diffusion Activation Visualizer |
|
|
|
|
|
Interactive visualization of diffusion model activations projected to 2D via UMAP. |
|
|
|
|
|
## Features |
|
|
- Explore activation space of diffusion models |
|
|
- Select points and find nearest neighbors |
|
|
- Generate images from averaged neighbor activations |
|
|
- Visualize denoising trajectories |
|
|
|
|
|
## Usage |
|
|
1. Hover over points to preview samples |
|
|
2. Click to select a point |
|
|
3. Click nearby points or use "Suggest KNN" to add neighbors |
|
|
4. Click "Generate from Neighbors" to create new images |
|
|
|
|
|
## Note |
|
|
First launch downloads ~2.5GB of data and checkpoints. Generation on CPU takes ~30-60s per image. |
|
|
|
