3D AI Zoomquilt Engine

A local, single-file desktop app that automates an infinite-zoom ("zoomquilt") video pipeline on top of a local Stable Diffusion WebUI, then lifts the flat frames into a 3D parallax fly-through and stitches an MP4 — all on your own GPU.

Tuned for an RTX 3060 12GB, but works on any CUDA card.

What it does

A four-pass pipeline, all local (the only network call is to your WebUI):

Pass	Stage	Output
1	2D outpaint loop — each frame shrinks the last and the model paints a new border	output/frames
2	Depth estimation (Depth-Anything-V2)	output/depths
3	3D camera warp (grid_sample radial parallax, depth-detrended to stay centered)	output/3d_warps
4	Zoom-tween interpolation + upscale + ffmpeg stitch	renders/zoomquilt_<timestamp>.mp4

Highlights:

• Dynamic timeline of prompt keyframes with backward-fill, a 🎲 dice that pulls from a 5,000-prompt landscape bank, and a settable keyframe interval.
• Small-gen → HQ-upscale: generate at a small fast resolution, upscale to 720p/1080p/1440p/4K with correct aspect (Cover/Fit) via ffmpeg lanczos, or a neural upscaler (R-ESRGAN via the WebUI).
• Smooth slow-motion via geometric frame interpolation (no extra generation), sub-pixel-accurate at the zoom center.
• Zoom In / Zoom Out direction, FPS, Speed, and a live "final resolution / length" readout.
• Fast-model controls — pick the checkpoint, steps, CFG, sampler from the app; one-click ⚡ Fast preset for Turbo/Lightning/Hyper-SD/LCM models.
• Backend health poll — auto-detects when your WebUI is online and auto-loads its styles/upscalers.
• Swappable prompt banks — a prompts/ folder of themed sets (landscapes, noir, space, trippy, creepy) in a dropdown; the 🎲 dice / Roll All draw from the chosen one. Drop in your own .txt and it appears automatically.
• Theme generator — type a theme ("+ New theme…" in the app, or make-theme.bat) and it writes a fresh prompt bank using your local Ollama (offline fallback included).
• Batch / Infinite — Bulk-N runs, ∞ Infinite, ⚄ Randomize-prompts-each-run, and a cooldown between runs, for an endless gallery of unique zoomquilts.
• Dark themed UI — collapsible, resizable panels (black / purple / orange).
• Multi-GPU aware (auto-pins the highest-VRAM card for the ML passes).

Requirements

• Python 3.11 (torch wheels; 3.13/3.14 are not yet supported by torch)
• A CUDA GPU
• ffmpeg on PATH (for the MP4 stitch)
• A local Stable Diffusion WebUI with the API enabled (see the installer below)

Quick start

:: 1. one-time setup (builds a Python 3.11 venv with CUDA torch + transformers)
setup.bat

:: 2. install an image-gen backend (pick one of 5) and launch it with the API on
install.bat

:: 3. run the app
launch.bat

In the app: Test Connection (it auto-detects when the backend is online) → set keyframe prompts → Generate Zoomquilt. The finished video lands in renders/.

Backend installer

install.bat opens a small GUI that git-clones and configures your choice of:

Backend	Plug-and-play*	Notes
SD WebUI Forge	⭐	Fast, low-VRAM, Flux/SDXL/SD1.5 — recommended
AUTOMATIC1111	⭐	The classic, broadest ecosystem
SD.Next	⭐	Performance-focused
Fooocus		Easiest SDXL; limited API
ComfyUI		Most flexible/fastest; native API (needs an adapter for this app)

* Plug-and-play = exposes the AUTOMATIC1111 /sdapi/v1/ API the app speaks, on port 7860.

For fast scenery generation, drop a few-step model (SDXL-Lightning, Hyper-SD, SD/SDXL-Turbo, or FLUX.1-schnell) into the backend's models folder and use the ⚡ Fast preset.

Files

File	Purpose
zoomquilt3d.py	The app (single file)
installer.py / install.bat	Backend installer & launcher
prompts/	Prompt banks (*.txt, one prompt per line) shown in the dropdown
build_prompt_bank.py	Regenerates prompts/landscapes.txt (5,000 prompts)
build_theme_banks.py	Regenerates the noir/space/trippy/creepy banks
theme_gen.py / make-theme.bat	Generate a new themed bank from a typed theme (Ollama)
setup.bat	Builds the Python 3.11 venv
launch.bat	Runs the app

License

MIT — see LICENSE.

---

Built iteratively with Claude Code.