🧊 Godmode 3D Creation: CLI-First & Local Engine 3D Pipeline

You are the Mastermind for all 3D generation, modeling, and CAD workflows within the BDB Creator OS. This skill operates on a CLI-First and Real-MCP paradigm, converting natural language prompts, concept images, or technical specifications into concrete terminal execution commands for local 3D engines, coupled with verified ecosystem MCP servers (Rhino, TouchDesigner, Unreal).

---

Plan first

Planning starts on the plan canvas with the research template, and the human approves it there — not in chat.

Choosing between engines, approaches or pipelines — evidence and recommendation reviewed before anyone builds.

aos-plan-canvas new research <plan-dir>                     # scaffold the plan folder
aos-plan-canvas open <plan-dir> --mode bdb-plan-builder  # render and open in the browser
aos-plan-canvas await <plan-dir>/plan.mdx                # keep listening until the verdict

The approved plan is the artifact: this template has no components, so there is no agenttrail map. The full loop and every template: the plan-canvas skill — a plain CLI, identical on every harness.

1. Engine & Local Backend Architecture

The core local 3D generation backends are hosted within the Creator Extension: bdb-dev-creator-extension/engines/3d/

The installer places the extension in its module directory (bdb-dev-creator-extension/ under the AOS module base). Every engines/3d/... path below is relative to that directory: cd there before running a command.

Primary Engines & Local Modules:

• TripoSR (engines/3d/triposr): Ultra-fast single-image 3D reconstruction for rapid mesh prototyping (under 0.5s on CUDA, CPU-capable).
• TRELLIS (engines/3d/trellis): High-fidelity Image-to-3D generation producing 3D Gaussian Splats, textured UV meshes, and radiance fields.
• Text-to-CAD (engines/3d/text-to-cad): Parametric engineering modeling using build123d and cadpy compiling to precise STEP, GLB, and STL files.

---

2. Concrete CLI Execution Specifications

When executing 3D workflows, run concrete terminal commands directly against the local Python engines.

A. TripoSR CLI (engines/3d/triposr)

Use run.py for instant single-image 3D mesh reconstruction for draft viewports, layout blocking, or real-time performance proxies.

# High-quality GLB export with baked texture:
python engines/3d/triposr/run.py <image_path> --output-dir <output_dir> --model-save-format glb --bake-texture

# Rapid mesh generation (OBJ with vertex colors, <0.5s on CUDA):
python engines/3d/triposr/run.py <image_path> --output-dir <output_dir> --mc-resolution 256

# Hardware Fallback: CPU mode (when CUDA GPU is unavailable):
python engines/3d/triposr/run.py <image_path> --output-dir <output_dir> --model-save-format glb --bake-texture --device cpu

B. TRELLIS CLI (engines/3d/trellis)

Use example.py for high-quality Image-to-3D generation producing PBR textured meshes, radiance fields, and Gaussian Splats.

# High-fidelity Image-to-3D GLB generation:
python engines/3d/trellis/example.py --image <image_path> --output <output_dir/model.glb>

# Direct pipeline execution with custom parameters:
python -c "from PIL import Image; from trellis.pipelines import TrellisImageTo3DPipeline; from trellis.utils import postprocessing_utils; pipeline = TrellisImageTo3DPipeline.from_pretrained('microsoft/TRELLIS-image-large').cuda(); outputs = pipeline.run(Image.open('<image_path>')); glb = postprocessing_utils.to_glb(outputs['gaussian'][0], outputs['mesh'][0], simplify=0.95, texture_size=2048); glb.export('<output_dir/model.glb>')"

C. Text-to-CAD CLI (engines/3d/text-to-cad)

Generate parametric CAD geometries and mechanical parts using build123d python scripts and the step compilation CLI.

# Step 1: Write parametric build123d script (<generator.py>):
# e.g., define geometry with build123d and export or pass to compiler

# Step 2: Compile generator script to STEP and GLB:
python engines/3d/text-to-cad/skills/cad/scripts/step <generator.py> -o <model.step> --glb <model.glb>

# Step 3: Inspect geometry facts and bounding planes (optional QA):
python engines/3d/text-to-cad/skills/cad/scripts/inspect refs <model.step> --facts --planes

---

3. Workflow Routing & Decision Flow

Select the appropriate engine and command pipeline based on input fidelity and hardware availability:

1. Rapid Real-Time Proxies / Draft Meshes (<1s):

   • Input: Concept sketch, photo, or transparent PNG.
   • Primary Command: TripoSR CLI (python engines/3d/triposr/run.py <image_path> --output-dir <output_dir> --model-save-format glb --bake-texture).
   • Output: Lightweight .glb or .obj for immediate stage placement.

2. High-Fidelity Visual Props & Hero Assets:

   • Input: Refined visual concept image.
   • Primary Command: TRELLIS CLI (python engines/3d/trellis/example.py --image <image_path> --output <output_dir/model.glb>).
   • Output: PBR-textured .glb with UV maps for Unreal Engine or TouchDesigner.

3. Parametric Engineering & Mechanical Parts:

   • Input: Dimensioned engineering constraints or geometric specifications.
   • Primary Command: Author Python build123d generator, then compile with python engines/3d/text-to-cad/skills/cad/scripts/step <generator.py> -o <model.step> --glb <model.glb>.
   • Output: Valid .step CAD model and companion .glb.

---

4. Hardware Constraints & Resilient Fallback Matrix

Execute directly via verified fallback mechanisms when hardware constraints or specific geometry requirements dictate:

1. CUDA Acceleration Unavailable (CPU-Only Environments):

   • TripoSR: Run with --device cpu: python engines/3d/triposr/run.py <image_path> --output-dir <output_dir> --model-save-format glb --bake-texture --device cpu
   • TRELLIS: TRELLIS requires CUDA for local inference. If CUDA is absent or VRAM is under 16GB, route to ComfyUI 3D diffusion workflow on port 8188 via comfyui-mcp (AOS ships no config for it: check it is configured in the harness first, as pb-image-to-3d does) or use an external cloud inference API endpoint.

2. Parametric CAD & Complex NURBS Modeling:

   • Primary Fallback: If build123d is unsuitable for complex freeform surfaces, use bdb_rhino_mcp_fallback (Rhino Grasshopper solvers, NURBS surfaces, curves, boolean operations, STEP/IGES export).

3. High-Fidelity Scene Assembly & Spatial Staging:

   • Use real ecosystem MCPs: bdb_unreal_mcp or bdb_td_minddesigner (create_3d_scene, create_point_cloud) to arrange, light, and render generated 3D assets in real time.

---

Overview

Godmode 3D Creation generates 3D assets, meshes, and CAD models using local Python CLI engines (TripoSR, TRELLIS, Text-to-CAD) and verified ecosystem MCP servers (Rhino, TouchDesigner, Unreal).

When to Use

• Trigger: The user asks to generate a 3D mesh, create a 3D model from an image, or construct a parametric CAD file.
• Exclude: Do not use for 2D video editing or live show timeline control.

Core Process

1. Identify the input type (sketch/photo for TripoSR, high-fidelity concept for TRELLIS, or dimensional specs for Text-to-CAD).
2. Check hardware availability (CUDA vs. CPU) and select the corresponding CLI invocation or fallback route.
3. Execute the CLI command and verify the generated .glb, .obj, or .step output file on disk.

Common Rationalizations

Rationalization	Reality
"I'll use TRELLIS for this quick draft."	TRELLIS requires substantial VRAM and execution time; use TripoSR for rapid mesh generation.
"I'll just guess the Text-to-CAD dimensions."	Parametric CAD requires strict units and tolerances; write precise build123d code with explicit millimeters.
"It previewed fine as an OBJ, we don't need textures."	High-fidelity assets require PBR textures and UV mapping; use --bake-texture with GLB export.
"CUDA is missing, so 3D generation is impossible."	TripoSR supports --device cpu, and CAD generation via build123d/step requires no GPU.

Red Flags

• Constructing fictitious JSON payloads for nonexistent MCP servers instead of running CLI commands.
• Failing to pass --device cpu when running TripoSR on non-CUDA systems.
• Passing vague natural language directly to Text-to-CAD without explicit millimeter dimensions in the generator script.
• Failing to check that the output asset file exists on disk after execution.

Verification

• CLI execution command targets the correct local engine script.
• Appropriate device mode (cuda or cpu) is specified.
• Output format (.glb, .obj, .step) matches the target pipeline.
• The generated 3D asset file exists in the target output directory.

---

Creator Pack Skills

<!-- aos:pack-index:start -->

• remotion — Generate walkthrough videos from Stitch projects using Remotion with smooth transitions, zooming, and text overlays
• threejs-skills — Create 3D scenes, interactive experiences and visual effects with Three.js. Use for 3D graphics, WebGL, animations or interactive 3D elements.
• spline-3d-integration — Use when adding interactive 3D scenes from Spline.design to web projects, including React embedding and runtime control API.
• bdb-after-effects-mcp — Utilizes the After Effects MCP servers (after-effects-mcp and ae-mcp) to create compositions, layers, masks, keyframe animations, and run ExtendScript.
• bdb-adobe-suite-mcp — >-
• bdb-davinci-mcp — Utilizes the DaVinci Resolve MCP servers to control timeline editing, media analysis, color grading, and Fusion/Fairlight scripting.
• bdb-blender-mcp — Utilizes the Blender MCP server and addon to build 3D assets, apply materials, inspect scenes, and script bpy.
• brag — >-
• design-spells — Use when adding curated micro-interactions, delightful animations, and subtle design details that add magic and personality to web and mobile apps.
• live-preview-canvas — A prototyping skill that generates a local HTML mock with a built-in feedback layer, allowing users to drop visual comments (red dots) for UI fixes.

Install: aos store install --pack creator

<!-- aos:pack-index:end -->