AGENT SKILL

ai3d-gen

Use when the user asks for generative 3D assets (text-to-3D, image-to-3D, mesh synthesis) via self-hosted open models. Routes to extensions/comfyui-operator/wrapper.py against a local ComfyUI server, with direct-model fallbacks (TRELLIS, TripoSG, InstantMesh, Hunyuan3D). All free/open weights, no paid SaaS.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 0
Forks
1

01 前言 Front Matter

name
ai3d-gen
description
Use when the user asks for generative 3D assets (text-to-3D, image-to-3D, mesh synthesis) via self-hosted open models. Routes to extensions/comfyui-operator/wrapper.py against a local ComfyUI server, with direct-model fallbacks (TRELLIS, TripoSG, InstantMesh, Hunyuan3D). All free/open weights, no paid SaaS.

03 SKILL.md 正文

AI 3D Generation (self-hosted)

Generate 3D assets locally via open-weight models. Two paths:

  1. ComfyUI-3D-Pack (primary): visual workflow server. Wrapper: extensions/comfyui-operator/wrapper.py (stdlib HTTP, no deps).
  2. Direct model CLIs (fallback / batch): run the model repos' own Python entrypoints, no ComfyUI needed.

Output of both paths: meshes (GLB/OBJ/PLY) that feed operate-blender for cleanup, retopo, texturing, render, or mesh-utils for inspection.

Prerequisites

  • ComfyUI installed locally (user step): git clone https://github.com/comfyanonymous/ComfyUI && pip install -r requirements.txt, then python main.py --port 8188.
  • ComfyUI-3D-Pack custom node pack for full 3D graph coverage: git clone https://github.com/MrForExample/ComfyUI-3D-Pack ComfyUI/custom_nodes/ComfyUI-3D-Pack && pip install -r ComfyUI/custom_nodes/ComfyUI-3D-Pack/requirements.txt
  • 3D-Pack requires NVIDIA GPU + CUDA toolkit (nvcc). It eagerly imports nvdiffrast, spconv-cu126, nerfacc, slangtorch; no CPU wheels exist. On CPU-only machines the whole pack fails import (ComfyUI core still runs). Use the direct-model fallback path there.
  • GPU strongly recommended for generation speed regardless (models are diffusion transformers; CPU inference works but is very slow).

Wrapper commands

python extensions/comfyui-operator/wrapper.py status          # health + VRAM
python extensions/comfyui-operator/wrapper.py submit wf.json  # queue job
python extensions/comfyui-operator/wrapper.py history <pid>   # poll status + outputs
python extensions/comfyui-operator/wrapper.py download <pid> --out dir/
python extensions/comfyui-operator/wrapper.py upload img.png  # input asset

submit requires API-format JSON (node-id -> {class_type, inputs}), the shape produced by ComfyUI's "Save (API Format)". Build it programmatically or export from the UI once.

history returns files: flattened output descriptors (filename, subfolder, type) regardless of which node emitted them. download fetches each via /view.

Model choices (open weights, license-verified in research)

Model Input Quality License Where
TRELLIS text/img high, GLB+texture MIT microsoft/TRELLIS
TripoSG img high mesh quality MIT VAST-AI/TripoSG
InstantMesh img fast Apache-2.0 TencentARC/InstantMesh
Hunyuan3D-2 text/img high + PBR Tencent permissive Tencent/Hunyuan3D-2
Stable Fast 3D img fastest UV+mat Stability Community stabilityai/stable-fast-3d

Prefer TRELLIS/TripoSG for permissive licensing; Hunyuan3D-2 when PBR texture quality matters; Stable Fast 3D for speed.

Intake preflight (photo-sourced pipelines)

Before spending a reconstruction/splat run on user photos:

  1. Provenance: source images should verify as real captures — run extensions/image-forensics/verdict.py per image when a camera reference exists; synthetic inputs break PRNU assumptions and usually produce bad splats.
  2. Photometric QC: run extensions/ppisp-prep/photo_qc.py on the set; correct/recapture verdicts route through the PPISP correction pass or a re-shoot before generation.

Recipes

image -> textured GLB (ComfyUI path)

  1. wrapper.py upload photo.png
  2. Build/submit a workflow JSON wiring LoadImage -> model node -> SaveMesh/Preview3D (node names per ComfyUI-3D-Pack version; query GET /object_info for exact class_type names on the live server).
  3. wrapper.py submit wf.json, poll history, download outputs.
  4. python extensions/mesh-utils/meshops.py info out.glb to verify.
  5. Import into Blender via operate-blender for retopo/material work.

text -> 3D without ComfyUI (direct)

  • Clone the model repo, pip install -r requirements.txt, run its inference script. E.g. TRELLIS example script produces GLB directly. Then mesh-utils/operate-blender downstream.

Boundaries

  • Never send user assets to hosted APIs (Meshy, Tripo cloud, etc.) unless explicitly asked. The point of this stack is self-hosted.
  • ComfyUI not running -> wrapper exits 1 with a start hint. Don't spawn the server silently; tell the user.
  • Check VRAM in status before queueing large models (>=8GB typical).

Model notes

  • ccore/Qwen3.5-4B-3D-Coder (ollama tag materion/qwen3.5-3d-coder:4b, ~4B params, ~2.7GB Q4_K_M) generates interactive three.js scenes in-browser from a prompt; a lightweight alternative to the ComfyUI pipeline for quick 3D prototypes.

04 同仓库更多 Skills

Leostruka
Leostruka/devin-bundle

montador-de-ideias

Use when assembling, expanding, or elevating a raw project idea or initial prompt into an ultra-detailed, first-principles dossier and compiled master prompt with journalistic OSINT market research, Tech Lead and Staff Architect justifications, atomic and modular hardening/optimization, and high-impact contact surfaces.

★ 0
Leostruka
Leostruka/devin-bundle

git-github-adaptive-engineering

Use when any task touches git history, GitHub issues/PRs/releases/CI, repo governance, or branch and merge decisions. Runs the Fase Zero situational-awareness battery, picks one of five delivery flows via the adaptive decision engine, and enforces commit/PR information-density standards. Orchestrates the gh, git-workflows, github-actions, pr, and debug/finish skills.

★ 0
Leostruka
Leostruka/devin-bundle

adaptive-agenda-calendar-routine-engine

Use when planning or executing work across hours, days, weeks, or sprints. Turns the calendar into a strict time-blocked agenda (HH:MM-HH:MM blocks carrying a routine profile, expected inputs, and a handoff output artifact), runs the 5-stage cognitive flow (Radar and Hunt -> Triage plus Steve Jobs conception -> workbench manifest -> >=32-cycle loop -> Close plus Open-Minded Scrum), and unfolds macro calendar milestones into micro agenda blocks through extensions/gcal/wrapper.py.

★ 0
Leostruka
Leostruka/devin-bundle

demo

Use when demonstrating a skill in a fixture project.

★ 0
Leostruka
Leostruka/devin-bundle

duplicate-skill

Use when testing duplicate skill name detection.

★ 0
Leostruka
Leostruka/devin-bundle

duplicate-skill

Use when testing duplicate skill name detection.

★ 0