Antoine Boucher

Make figurines with AI for 3D printing

Good 2D toy art beats a fancy image-to-3D model. This post walks the steps in order, links a slide deck, a ComfyUI workflow, and how I sketch the flow with uml-mcp.

Pipeline (in order)

  1. Reference photo — clear side or front view.
  2. Chibi vinyl render — ChatGPT, Midjourney, or ComfyUI (OmniGen2).
  3. Mesh — Hunyuan3D v2 (local ComfyUI or hosted tools).
  4. Print — orient, supports, resin slice; fix islands with UVtools if needed.

Photo to Object — photo, AI toy image, 3D model, resin print

Pipeline sequence (source: figurine-pipeline-sequence.mmd, rendered with uml-mcp):

Figurine pipeline — reference photo to resin print

sequenceDiagram
    autonumber
    actor User
    participant Photo as Reference photo
    participant ComfyUI
    participant OmniGen2 as OmniGen2 stage
    participant Hunyuan3D as Hunyuan3D v2 stage
    participant Slicer as Resin slicer
    participant Printer

    User->>Photo: Side or front reference
    User->>ComfyUI: Load chibi workflow + reference-photo.jpg
    ComfyUI->>OmniGen2: ReferenceLatent + chibi vinyl prompt
    OmniGen2-->>ComfyUI: PNG (white background)
    ComfyUI->>Hunyuan3D: CLIPVision + image-to-mesh
    Hunyuan3D-->>ComfyUI: GLB (VoxelToMesh 0.7 / 0.6)
    ComfyUI-->>User: mesh/ComfyUI*.glb
    User->>Slicer: Orient, supports, hollow
    Slicer-->>User: Sliced job
    User->>Printer: Resin print
    Printer-->>User: Chibi figurine

Deck (~17 MB) — screenshots and tool URLs from the talk:

Download slide deck (.pptx)(~17 MB)

In-browser preview via Microsoft Office Online. If it fails to load, use the download link above (large decks may not preview).

Title slide — Make figurines using AI for 3D printing

Step 1 — 2D toy / chibi

ChatGPT or Midjourney for a toy plastic figure

Deck prompt:

Convert this image into a toy-style plastic figure. Keep features but simplify and smoothen as if molded in plastic.

Before / after Toy result

Chibi prompt (same idea, bigger head):

Vinyl chibi figurine, smooth plastic, white background, one character, pose that can stand on a base.

Gallery (ChatGPT):

Bernese puppy Golden retriever Cowboy Pool Spider

Document the flow — uml-mcp

The sequence diagram above was generated with uml-mcp (generate_uml, type mermaid) and saved as figurine-pipeline-sequence.mmd + SVG in this page bundle.

Chat → uml-mcp MCP → diagram

In Cursor: enable the uml-mcp server (repo README), then call generate_uml with the .mmd source or ask for a sequence diagram of the ComfyUI → print steps. Same idea as ChatGPT sequence diagrams.

Step 2 — ComfyUI (one graph)

Download: image_omnigen2_image_edit.json — Load in ComfyUI (OmniGen2 chibi edit + Hunyuan3D v2 mesh). Put your photo in input/ as reference-photo.jpg (or edit the LoadImage node).

StageWhat it does
AOmniGen2 + ReferenceLatent → chibi PNG (heun, 45 steps)
BHunyuan3D v2 → two GLBs (mesh/ComfyUI*.glb, thresholds 0.7 and 0.6)

Full positive/negative prompts and seeds are in the JSON. Needs OmniGen2 + Hunyuan3D-2 custom nodes (Hunyuan3D-2). On 8–12 GB VRAM, lower octree_resolution or num_chunks.

Step 3 — Image to 3D (hosted)

Convert to 3D with Hunyuan3D and Rodin

Single-image mesh tools — clarity of the toy PNG matters more than the brand.

ToolLinkNotes
Hunyuan3D-2GitHubOpen weights / local or hosted runs; strong general image-to-3D.
RodinHDWebGL demoHigh-detail avatars and figurines from images.
Hi3DGenHugging Face demo2D diffusion prior + geometry consistency (CHI 2024 paper, white paper PDF).
Hyper3Dhyper3d.aiFast commercial image-to-3D.
Meshymeshy.aiText/image to 3D with editing and remeshing.
Tripo3Dstudio.tripo3d.aiBrowser studio workflow.
Sparc3Dlizhihao6.github.io/Sparc3DStructured sparse-view reconstruction.

Tripo3D studio

Hyper3D

Sparc3D

Meshy AI

Hi3DGen (CHI 2024, demo) — strong when the 2D silhouette is already clean.

Mesh preview

Sample GLB files from Hunyuan3D, RodinHD, Hyper3D, and Hi3DGen are not stored in this repo (each was 20–75 MB). Use the generator previews in the tool UIs, or export from your own ComfyUI run (mesh/ComfyUI*.glb).

Mesh preview in generator UI

Step 4 — Mesh settings

Generator settings (animated)

Mesh preview (animated)

Settings detail

Settings panel

Tune face count, symmetry, and remesh in the generator UI before export.

Step 5 — Resin print

Resin print setup

Printed figurine on build plate

Resin for small parts. Tilt overhangs toward the plate; light supports on hair and capes.

Islands in slicer

Check floating islands with the slicer or UVtools before printing.

Printed results

Finished figurine 1

Finished figurine 2

When to use this pipeline

GoalStart here
Gift figurine from a pet photoChatGPT/Midjourney chibi → hosted mesh
Repeatable local batchComfyUI JSON in this bundle
Teaching a classDeck + uml-mcp sequence diagram
Production CAD interchangeExpect cleanup in Blender after GLB

Pitfalls

  • Skipping 2D quality — mesh tools amplify bad silhouettes.
  • Committing 50 MB GLBs — export locally; link tools instead.
  • Zero island check — resin failures mid-print waste hours.

References