Neural image synthesizer — text-to-image and animation powered by CLIP + VQGAN/Limited Palette

PyTTI Portable is a self-contained distributable of pytti-core with a Gradio web UI. Everything is bootstrapped from a single install.bat — no system-wide Python required.

• Gradio web UI with live preview, config save/load, and render controls
• 3D animation with camera transforms (translate, rotate, zoom) and depth estimation
• Multiple image models: Limited Palette, VQGAN (multiple checkpoints)
• CLIP-guided rendering with multi-model ensemble (ViT-B/32, ViT-B/16, RN50x4, etc.)
• Video Source mode for style transfer onto existing video
• Breath mode — linear crossfade from init image to CLIP-optimized output over the full render. Frame 1 is your original image; the final frame is fully transformed by CLIP. Creates smooth "emergence" animations showing the AI interpretation gradually taking over
• Video encoding — built-in ffmpeg encoding (MP4 H.264, ProRes 4444, ProRes HQ) directly from the Output tab
• Audioreactive animation support
• Config system powered by Hydra — save, load, and share render presets as YAML files

• Windows 10 or 11
• NVIDIA GPU (GTX 10xx through RTX 40xx)
  • RTX 50xx (Blackwell) is not supported
• Up-to-date NVIDIA drivers
• Git installed and on PATH
• ~8 GB disk space (after install)

git clone https://github.com/pxl-pshr/pytti.git
cd pytti

1. Double-click install.bat (one-time, ~30-60 min)
2. Double-click launch.bat
3. A browser window opens — start rendering

To update: git pull from the pytti folder.

The first render will download CLIP and depth models (~1-4 GB), cached after that.

pytti/
├── install.bat          # One-time installer (downloads Python, PyTorch, deps)
├── launch.bat           # Starts the Gradio UI
├── app/
│   ├── ui.py            # Gradio web UI
│   ├── patch_gradio.py  # Post-install dependency patches
│   └── config/
│       ├── default.yaml # Default render settings
│       └── conf/        # User-saved presets
└── examples/            # Sample renders

PyTTI uses CLIP to guide an image generator (Limited Palette or VQGAN) toward text prompts. In animation mode, each frame is warped via 2D/3D transforms with AdaBins depth estimation, then re-optimized toward the prompt — producing dreamlike, evolving visuals.

• Demo video — pytti in action
• pytti-book — documentation, tutorials, and parameter guide
• pytti-motion-preview — camera motion preview tool
• pytti-notebook — the original Colab notebook this project is based on

• David Marx & sportsracer48 — original pytti creators and maintainers
• pytti-core — the rendering engine
• CLIP — OpenAI's vision-language model
• taming-transformers — VQGAN
• AdaBins — monocular depth estimation (3D mode)
• GMA — optical flow for video mode
• Gradio — web UI framework

MIT — see LICENSE for details.