POD AI Reality Lab · AIFENG DCS

❓ The Boss's Question

Is Lingtu really a pure paid-API in-house model? Or is it 'open-source models + self-built fixed code'? This lab rebuilds POD AI for the digital printing industry along the 'open-source models + self-built workflow' route: transparent cost, data never leaves the machine, and capability that can prove itself.

🖥 This Machine's Hardware (everything runs locally, data stays inside)

🎮 GPU: NVIDIA RTX 5080 Laptop (16 GB VRAM)

🧠 Memory: 65 GB

⚙ CPU: 24 threads

16 GB VRAM is enough for SDXL fp8 (≈7 GB) and FLUX.1-dev fp8 (≈12 GB); the prepress pipeline runs entirely on CPU and never competes with GPU generation.

Three-Layer Architecture Overview

L1 · Fixed-code prepress pipeline (Python scripts, no AI inference) Function share: ≈70% of functions Cost: ≈¥0 (electricity only)

Included capabilities：rembg background removal · Real-ESRGAN HD upscaling · vtracer vectorization · OpenCV DPI check + bleed + safe area · Pillow text layout / color pick / recolor

70% of daily digital-printing work is deterministic image processing — fixed code is more stable, faster, and free compared with AI.

L2 · Local ComfyUI open-source generation (GPU inference) Function share: ≈25% of functions Cost: ¥0 (electricity only)

Included capabilities：SDXL base model · FLUX.1-dev fp8 · ControlNet line-art control · IPAdapter reference-style transfer (print-pattern extraction) · LoRA industry styles

Creative images that truly need 'generation' are only about a quarter — all handled by local open-source models with quality comparable to paid services.

L3 · Paid-API fallback (only when unavoidable) Function share: ≤5% of scenarios Cost: Pay per image

Included capabilities：Only for hard cases like complex Indonesian/English text posters — via Ideogram (local models render text poorly).

Money is spent only on the rare cases local models genuinely cannot do, and every call is recorded in the audit log.

🔍 Honesty Statement

We cannot verify Lingtu's internal implementation (a black-box service) — we state this plainly. But the industry-standard practice = open-source models + a self-built workflow wrapper. This plan does not rely on guessing Lingtu; it builds equivalent capability on a public open-source stack and proves it via blind testing (see the Blind Test page).

This architecture diagram is also the construction blueprint: L1 first (pure scripts, zero risk), L2 follows, L3 last.

PENDING · awaiting boss approval

⚠ The boss is asked again before every single step starts