herbygitea/pbs-projects
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
27f43a88e0
pbs-projects/Tech/Projects/instantmesh-docker.md

2.7 KiB
Raw Blame History

project type status path tags created updated
instantmesh-docker project-plan active Tech/Projects
homelab
docker
ai-ml
gpu
2026-04-27 2026-04-27

InstantMesh Docker Deployment

Self-hosted deployment of TencentARC/InstantMesh on the Manjaro tower (RTX 3080) for single-image to 3D mesh generation. Gradio UI exposed on port 43839.

Goals

  • Run InstantMesh locally via Docker Compose on the Manjaro tower
  • Persist downloaded model weights outside the container (re-pulls are slow)
  • Build from official repo source, patch the known onnxruntime issue from upstream

Stack

  • Host: Manjaro tower (i7 / 128GB / RTX 3080, 10GB VRAM)
  • Runtime: Docker + NVIDIA Container Toolkit
  • Base image: Built from TencentARC/InstantMesh/docker/Dockerfile
  • UI: Gradio on :43839
  • Model variant: instant-mesh-large (default, fits in 10GB VRAM)

Known issue

Upstream Dockerfile has a missing dep — container fails at runtime with ModuleNotFoundError: No module named 'onnxruntime' (issue #175). Fix is to add onnxruntime to the pip install step in the Dockerfile before building, or install it post-build.

Setup steps

  • Verify NVIDIA Container Toolkit is functional: docker run --rm --gpus all nvidia/cuda:12.1.0-base-ubuntu22.04 nvidia-smi
  • Clone repo: git clone https://github.com/TencentARC/InstantMesh.git ~/projects/instantmesh
  • Patch Dockerfile to add onnxruntime to pip install
  • Build image: docker compose build
  • Create model dir: mkdir -p ~/instantmesh-models
  • First run: docker compose up — model weights download on first launch (~10GB)
  • Verify Gradio UI at http://localhost:43839
  • Test inference with sample image

docker-compose.yml

yaml services: instantmesh: build: context: ./InstantMesh dockerfile: docker/Dockerfile image: instantmesh:local container_name: instantmesh ports: - "43839:43839" volumes: - ~/instantmesh-models:/workspace/instantmesh/models deploy: resources: reservations: devices: - driver: nvidia count: all capabilities: [gpu] restart: unless-stopped

Open questions

  • Reverse proxy through Traefik on *.lab.herbylab.dev, or keep local-only?
  • Worth wiring into OpenClaw later as a callable 3D-gen tool?

References

...sent from Jenny & Travis