Initial seed upload

SPACES_README.md

@@ -0,0 +1,30 @@
+---
+title: DiffViews
+emoji: 🔬
+colorFrom: purple
+colorTo: blue
+sdk: gradio
+sdk_version: 4.44.0
+app_file: app.py
+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.

app.py

@@ -0,0 +1,164 @@
+"""
+HuggingFace Spaces entry point for diffviews.
+
+This file is the main entry point for HF Spaces deployment.
+It downloads required data and checkpoints on startup, then launches the Gradio app.
+
+Environment variables:
+    DIFFVIEWS_DATA_DIR: Override data directory (default: data)
+    DIFFVIEWS_CHECKPOINT: Which checkpoint to download (dmd2, edm, all, none; default: dmd2)
+    DIFFVIEWS_DEVICE: Override device (cuda, mps, cpu; auto-detected if not set)
+"""
+
+import os
+from pathlib import Path
+
+# Data source configuration
+DATA_REPO_ID = "mckell/diffviews_demo_data"
+CHECKPOINT_URLS = {
+    "dmd2": (
+        "https://huggingface.co/mckell/diffviews-dmd2-checkpoint/"
+        "resolve/main/dmd2-imagenet-64-10step.pkl"
+    ),
+    "edm": (
+        "https://nvlabs-fi-cdn.nvidia.com/edm/pretrained/"
+        "edm-imagenet-64x64-cond-adm.pkl"
+    ),
+}
+CHECKPOINT_FILENAMES = {
+    "dmd2": "dmd2-imagenet-64-10step.pkl",
+    "edm": "edm-imagenet-64x64-cond-adm.pkl",
+}
+
+
+def download_data(output_dir: Path) -> None:
+    """Download data from HuggingFace Hub."""
+    from huggingface_hub import snapshot_download
+
+    print(f"Downloading data from {DATA_REPO_ID}...")
+    print(f"Output directory: {output_dir.absolute()}")
+
+    snapshot_download(
+        repo_id=DATA_REPO_ID,
+        repo_type="dataset",
+        local_dir=output_dir,
+        revision="main",
+    )
+    print(f"Data downloaded to {output_dir}")
+
+
+def download_checkpoint(output_dir: Path, model: str) -> None:
+    """Download model checkpoint."""
+    import urllib.request
+
+    if model not in CHECKPOINT_URLS:
+        print(f"Unknown model: {model}")
+        return
+
+    ckpt_dir = output_dir / model / "checkpoints"
+    ckpt_dir.mkdir(parents=True, exist_ok=True)
+
+    filename = CHECKPOINT_FILENAMES[model]
+    filepath = ckpt_dir / filename
+
+    if filepath.exists():
+        print(f"Checkpoint exists: {filepath}")
+        return
+
+    url = CHECKPOINT_URLS[model]
+    print(f"Downloading {model} checkpoint (~1GB)...")
+    print(f"  URL: {url}")
+
+    try:
+        urllib.request.urlretrieve(url, filepath)
+        print(f"  Done ({filepath.stat().st_size / 1e6:.1f} MB)")
+    except Exception as e:
+        print(f"  Error downloading checkpoint: {e}")
+        print("  Generation will be disabled without checkpoint")
+
+
+def ensure_data_ready(data_dir: Path, checkpoints: list) -> bool:
+    """Ensure data and checkpoints are downloaded."""
+    # Check if data exists (look for config files)
+    has_data = any(
+        (data_dir / model / "config.json").exists()
+        for model in ["dmd2", "edm"]
+    )
+
+    if not has_data:
+        print("Data not found, downloading...")
+        download_data(data_dir)
+    else:
+        print(f"Data found in {data_dir}")
+
+    # Download checkpoints
+    for model in checkpoints:
+        download_checkpoint(data_dir, model)
+
+    return True
+
+
+def get_device() -> str:
+    """Auto-detect best available device."""
+    override = os.environ.get("DIFFVIEWS_DEVICE")
+    if override:
+        return override
+
+    import torch
+
+    if torch.cuda.is_available():
+        return "cuda"
+    if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+        return "mps"
+    return "cpu"
+
+
+def main():
+    """Main entry point for HF Spaces."""
+    # Configuration from environment
+    data_dir = Path(os.environ.get("DIFFVIEWS_DATA_DIR", "data"))
+    checkpoint_config = os.environ.get("DIFFVIEWS_CHECKPOINT", "dmd2")
+    device = get_device()
+
+    # Parse checkpoint config
+    if checkpoint_config == "all":
+        checkpoints = list(CHECKPOINT_URLS.keys())
+    elif checkpoint_config == "none":
+        checkpoints = []
+    else:
+        checkpoints = [c.strip() for c in checkpoint_config.split(",") if c.strip()]
+
+    print("=" * 50)
+    print("DiffViews - Diffusion Activation Visualizer")
+    print("=" * 50)
+    print(f"Data directory: {data_dir.absolute()}")
+    print(f"Device: {device}")
+    print(f"Checkpoints: {checkpoints}")
+    print("=" * 50)
+
+    # Ensure data is ready
+    ensure_data_ready(data_dir, checkpoints)
+
+    # Import and launch visualizer
+    from diffviews.visualization.app import GradioVisualizer, create_gradio_app
+
+    print("\nInitializing visualizer...")
+    visualizer = GradioVisualizer(
+        data_dir=data_dir,
+        device=device,
+    )
+
+    print("Creating Gradio app...")
+    app = create_gradio_app(visualizer)
+
+    print("Launching...")
+    # HF Spaces expects server on 0.0.0.0:7860
+    app.queue(max_size=20).launch(
+        server_name="0.0.0.0",
+        server_port=7860,
+        share=False,  # Spaces handles public URL
+    )
+
+
+if __name__ == "__main__":
+    main()

requirements.txt

@@ -0,0 +1,23 @@
+# DiffViews - HuggingFace Spaces Requirements
+# Install diffviews package from GitHub
+git+https://github.com/mckellcarter/diffviews.git
+
+# Core dependencies
+torch>=2.0.0
+numpy>=1.21.0
+pandas>=1.5.0
+pillow>=9.0.0
+scikit-learn>=1.0.0
+umap-learn>=0.5.0
+tqdm>=4.60.0
+
+# Visualization
+gradio>=4.0.0
+plotly>=5.18.0
+matplotlib>=3.5.0
+
+# HuggingFace Hub for data download
+huggingface_hub>=0.19.0
+
+# Optional but useful
+scipy>=1.7.0