I wanted to share a project I’ve been building over at Grumpy's Workshop (McKeand Consulting, LLC) that I think will be a game-changer for anyone looking to create deep, organic 3D carvings without spending a fortune on expensive software models or specialized design suites.
It is called the Local AI 2.5D Relief Engine, and it is completely open-source (MIT License). I am looking for a handful of community members to help me test it out, look over the workflow, and break things before a wide release.
🧭 Where the Project Came From
The idea started because I wanted a smooth, predictable way to generate high-quality 2.5D relief files natively for Vectric VCarve. I spent quite a bit of time experimenting with things like ComfyUI, InvokeAI, and SD.Next, but honestly, it felt like a hot mess. The node trees were overly complicated, custom nodes were needed, the hardware limits were frustrating, and getting a clean, 16-bit displacement map that didn’t turn into a jagged, stepped mess or a featurless blob on a carving machine was incredibly difficult.
Vectric recently introduced their own tools toward this (EasyCreate), but I wanted something specialized and running 100% locally on my workstation without external cloud fees or internet lag.
🪄 Vibe Coding, Antigravity, and My Sandbox Setup
I didn't write this complex pipeline alone. I built this entire engine using a concept called "Vibe Coding". By teaming up with an advanced AI collaborator and utilizing Antigravity—a highly specialized agentic development framework—we co-developed the entire codebase in record time.
Instead of getting bogged down in syntax errors, I acted as the systems architect and woodcarver. I described the physical math involved, the toolpath problems (like steep vertical plunges or recessed nose/eye shadows), and the AI generated the code. I engineered and optimized this system directly on a local AI Workstation environment:
- The System: A repurposed a Dell XPS-8920 with a i7-7700, 64 GB RAM, a 1TB SSD for root, a 2TB HD used for output storage (for image generation for example), Wayland is running on the intel HD Graphics 630 and a NVIDIA RTX 5060 Ti OC with 16 GB VRAM for AI tasks. (Old - except for the GPU - I know but I didn’t want to spend a fortune on this – the RAM and GPU were purchased before prices soared.)
- The OS: The system is running Ubuntu Studio 26.04.
- The Accelerator: Local GPU execution driven by a high-performance NVIDIA RTX 5060 Ti.
🛠️ What the Engine Actually Does
The dashboard runs a local web UI right in your browser (built on Python and Streamlit – a framework for interactive web applications). It does two big things:
- AI Art & Advanced Depth Estimation: You type in a simple subject (like "a bugling elk"), and an integrated local LLM automatically enriches the prompt with woodcarving properties. It synthesizes an image and uses a powerhouse model called Marigold to estimate true 3D spatial contours—ensuring parts like eyes and noses dome outward instead of sinking inward like normal light/shadow pictures.
- 12-Stage 16-Bit CNC Preprocessing Pipeline: It instantly maps your designs into a lossless 16-bit integer array (65,536 height levels). This prevents that horrible "stair-step" terracing effect on slopes.
The CNC specific controls include:
- Background Clamp & Corner Flood-Fill: Automatically zeroes out connected background vignettes so your pocket floor stays perfectly flat.
- Shadow Lift: Gradually raises ultra-deep interior crevices to protect fine carving bits from plunging too steeply.
- Highlight Apex Crowning: Detects flat plateau areas near peak height and rounds them into convex domes.
- Watertight STL Generation: Outputs a solid 300mm x 300mm reference block with a 19mm (0.75") maximum carve depth and a 3mm safe bottom buffer. You can export as a solid STL mesh or a raw 16-bit grayscale .png heightmap.
- Although we specify a reference size as 300mm x 300mm x (19 mm + 3 mm) the STLs are actually unitless. When you import the STL into VCarve you will need to use the “Scale mm / inch” to bring the mesh down to a usable scale.
⚠️ Crucial Limitations & Commercial Boundaries (Please Read!)
While the orchestration source code I wrote is released under the permissive MIT License, the underlying AI models downloaded at runtime (specifically the Marigold 3D depth and surface normal pipelines) are governed by strict upstream CC BY-NC-SA 4.0 (Non-Commercial) terms. Because of this, we have very definitive legal boundaries:
- No Online/SaaS Platforms: You cannot use this project's code or components to spin up your own public online image-to-relief subscription platform, website, or paid remote service.
- Physical Carvings for Profit = ✅ YES: You are 100% permitted to take the STL files or heightmaps you generate, carve them into physical wood or composite material blocks on your CNC, and sell those physical carvings commercially.
- Selling the Digital STLs = ❌ NO: You cannot sell or commercially distribute the digital STL meshes, OBJ files, or 16-bit heightmap image files generated by the Marigold pipeline. They are strictly for your own personal machine use.
(Note: If you bypass the AI depth estimation entirely and use the built-in "Direct Luminance" mode with your own commercially permissible or public domain 2D assets, these non-commercial restrictions do not apply).
👥 How to Join the Testing Pool (Zero Friction)
I know navigating Git repositories can be annoying, and I want to make sure you do NOT need a GitHub account to participate.
I am hosting a secure, private instance of the platform right out of my business office environment using a lightweight platform similar to GitHub called Gitea (Git with a cup of tea).
- If you want to be a tester, just reply to this thread or send me an email!
- I will manually spin up a dedicated tester account for you on my secure server.
- You can log in, grab the code or instructions, use the built-in issue tracker to report bugs, and share your carving results.
🎬 Demo Video for the Relief Engine