PAPERPLANES

Slicing the two-dimensional lie into a stratified hallucination.

PaperPlanes is a sequential pipeline designed to dismantle the static nature of 2D paintings and reconstruct them as stratified, multi-layered 3D paper theater layers. This is not a "filter"; it is a multimodal reconstruction and spatial reconciliation engine.

---

I. THE ARCHITECTURE

1. The Wizard (Frontend PWA)

A monochromatic, "Glitch-Noir" interface that manages the pipeline’s state.

• Object Permanence: Tracks unique run_id state changes to prevent the ingestion of stale artifacts from the GitHub cache.
• Manual Overrides: Provides vertex-dragging for perspective warping, text-injection for semantic reconstruction, and a 5-axis affine transformation suite for depth reconciliation.

2. The Smuggler (Cloudflare Worker Proxy)

A serverless bridge that bypasses GitHub’s aggressive 5-minute CDN cache.

• Binary Reconstruction: Fetches file metadata, retrieves raw Base64 strings, and decodes them in-flight into a Uint8Array to serve fresh image bytes directly to the Wizard.

3. The Executioner (GitHub Action Runner)

The compute-heavy viscera of the operation.

• Serial Sanity: Uses concurrency groups to ensure a single lane of execution, preventing repository corruption during simultaneous triggers.
• Multimodal Engine: Leverages Gemini 1.5 Flash to deconstruct and reconstruct imagery, and Depth-Anything-V2 for spatial inference.

---

II. THE PIPELINE

1. CROP: Auto-detection or manual 4-point vector warp. Result: cropped_image.png.
2. GENERATE: Multimodal reconstruction. Painting becomes photograph. Result: generated_image.png.
3. DEPTH: Neural spatial inference. Result: raw_depth_map.png.
4. ALIGN: Human-in-the-loop affine sync. Result: realigned_depth_map.png.
5. SEGMENT: Linear binning and mask smoothing. Result: paper_planes_layers.zip.

---