1. Pick the tools
For parts with real dimensions, a script beats a picture: you change one number and re-run it. All of these run from the command line or a Python script, with no window open. That is what "headless" means in the table.
| Tool | Licence | Headless use |
|---|---|---|
| OpenSCAD | GPL-2.0 with a CGAL linking exception [1] | Its own scripting language. openscad -o file.stl model.scad runs without the GUI and -D var=val overrides parameters. Exports include STL, 3MF and SVG [1]. |
| CadQuery | Apache-2.0 | Python library on the OpenCascade kernel. Exports STEP, STL, 3MF and others [2]. |
| build123d | Apache-2.0 | Python library on the OpenCascade kernel. export_stl() and export_step(); 3MF through its Mesher class [3]. |
| FreeCAD | LGPL-2.1 | FreeCADCmd gives a console without the GUI and runs scripts. Some features need the GUI under a virtual display [4]. |
| Blender (bpy) | GPL-3.0 | blender -b runs without the UI; bpy is also a Python module [5]. Useful for rendering previews, not needed to print. |
| trimesh (mesh repair) | MIT | Python. trimesh.repair can fill holes and fix normals and winding. Test the repair on your own mesh [6]. |
| PyMeshLab (mesh repair) | GPL-3.0 | Python bindings for MeshLab. We could not confirm from its pages whether it runs fully headless [7]. |
| OrcaSlicer (slicer) | Not checked by us | The official wiki says it can load, transform and slice from the command line without the GUI [8]. |
| PrusaSlicer (slicer) | Not checked by us | The command line has --export-gcode, --export-3mf and --load for a config file [9]. |
Our suggestion: start with OpenSCAD or build123d for the part and a slicer command line for the check. Slicer time and filament numbers are only as good as the printer profile you give them, and estimates can differ from real print time. We have not tested whether the slicer command lines run without a display on a pod.
2. AI 3D generators, and do you need a GPU pod?
These open models turn text or a photo into a mesh. Memory figures are the minimums each project states, and the licences are what each page says. Read the licence on the model's page yourself before you use it.
| Model | Licence | Stated GPU memory |
|---|---|---|
| TripoSR | MIT [13] | About 6 GB for one image at default settings. Needs a CUDA-built torchmcubes for mesh extraction [13]. |
| Stable Fast 3D | Stability AI Community Licence: free for commercial use up to US$1M annual revenue, an enterprise licence above that. The download is gated: you accept conditions and log in [14]. | About 6 GB for one image. A CPU fallback exists (SF3D_USE_CPU=1); we do not know how slow it is [14]. |
| TRELLIS | MIT for the models and most of the code; some submodules differ [10] | At least 16 GB. Verified by the project on A100 and A6000, Linux only [10]. |
| TRELLIS.2 | MIT for the model and code; some dependencies carry their own licences [11] | At least 24 GB. Verified by the project on A100 and H100 only; a 24 GB RTX 3090 or 4090 meets the stated minimum but was not among the tested cards [11]. |
| Hunyuan3D-2.1 | Tencent community licence. It excludes the European Union, the United Kingdom and South Korea; above 1 million monthly active users you must request a licence from Tencent; outputs may not be used to improve other AI models [12]. | About 10 GB for shape, 21 GB for texture, 29 GB for shape plus texture. A --low_vram_mode flag exists; we do not know how much it helps [12]. |
- No GPU needed for OpenSCAD, CadQuery, build123d, FreeCAD headless, mesh repair and slicer command lines.
- A GPU pod helps for TRELLIS, TRELLIS.2 and Hunyuan3D-2.1, whose READMEs target CUDA (NVIDIA) GPUs [10][11][12]. A 24 GB card covers TripoSR, Stable Fast 3D, TRELLIS and the Hunyuan shape step. Shape plus texture in Hunyuan needs about 29 GB, so plan a 48 GB card [22].
- A CPU pod exists for the code-CAD route if your own device cannot run the tools, for example an iPad. The RunPod console shows the exact price before you deploy, so we do not quote one here.
Example cost, rates checked within the last 30 days, billed by the second: an RTX 3090 (24 GB) is $0.50 per hour on Secure Cloud and $0.22 on Community Cloud. An RTX A6000 (48 GB) is $0.69 per hour Secure and $0.33 Community [21]. An hour on a 3090 is therefore $0.50 on Secure Cloud. That is arithmetic, not a measurement, and we have not generated models on a pod ourselves yet.
3. What AI is good and bad at for printing
This section is our judgement, not a tested result. No source we read checks these claims, so read them as cautions.
- Usually a good fit: parametric parts (brackets, stands, organizers and enclosures), where the assistant writes a script with named dimensions and you can re-run it with new numbers. Decorative models and figurines from AI generators also fit, if you check them first.
- Be careful with wall thickness and overhangs. The AI generators we read about do not state any check for either. Ask for a printability check and look at the slicer preview.
- Be careful with tolerances and fits. The assistant does not know your printer. Print a small test piece before you print the whole part.
- Scale is arbitrary. An AI-generated mesh has no real-world size until you scale it, and then you must measure it. Until then it could be a keyring or a garden shed. The file does not say.
- Do not rely on AI for load-bearing or safety-critical parts. An assistant cannot certify strength. See section 7.
We found no independent benchmark of how often these models' outputs pass slicer checks. Expect to check them, and sometimes repair them.
4. From zero to a first printable file
- Decide if you need a pod. A bracket, stand or organizer is a code-CAD job: use your own laptop if it can run the tools. Use a GPU pod only for AI mesh generation.
- Collect what your assistant needs: printer model, nozzle, material, build volume, your measurements in millimetres, and how the part will be used. A photo of the thing it must fit helps.
- If you do want a pod, have your assistant quote the hourly price and wait for your yes, or deploy it yourself in the RunPod console.
- Have the assistant install the tools and print each tool's version.
- Have it write a script with named dimensions, export STL or 3MF, and report the bounding box (the overall size) in millimetres, the volume, and whether the mesh is watertight (fully closed, with no holes).
- Look at the part from several angles, either in rendered images or in a viewer on your own device. Never print a file you have not looked at.
- Slice it with your printer, material and layer height, and read the time, filament use and warnings. Then always preview the layers in a slicer before printing: missing walls, thin features and long unsupported spans show up there.
- Download the files to your device, then terminate the pod and check the Pods page that nothing is running.
- At home, slice again in your own slicer, check the layer preview, and print a small or scaled-down test first if fit matters.
5. If you use a pod: what to set up
- Account: you create the account and add a small balance yourself, such as $10. Do not hand your AI assistant your RunPod key or payment details unless you accept that it can spend money. Create a RunPod account (referral link: we may earn credits if you sign up and use RunPod). What happens when you press it.
- Templates: for code CAD, the official CPU templates "Runpod Ubuntu 22.04" or "Runpod Ubuntu 24.04". For AI generation, the official "Runpod Pytorch 2.8.0" template (CUDA 12.8, JupyterLab and SSH on) [15]. We found no template with OpenSCAD, CadQuery, build123d, Blender or a slicer preinstalled, so your assistant installs them after the pod starts.
- Where installs live: the container disk is erased when a pod stops, while
/workspace(the volume disk) is kept [15][20]. - Getting files out:
runpodctl sendon the pod prints a one-time code, andrunpodctl receiveon your computer fetches the file. It needs a computer that can run runpodctl, so it does not suit an iPad [18].scpworks over SSH with a public IP and TCP port, but not over the basic proxied SSH method [19]. You can also use the JupyterLab file browser. Upload by drag and drop is documented; we have not tested download, including in Safari on an iPad [16]. RunPod charges no fees for data in or out [21]. - Web addresses: a pod can publish
https://POD-ID-PORT.proxy.runpod.net. That address is public with no login of its own, and requests through it time out after 100 seconds [17]. Do not leave a file server or model viewer running on it; your designs would be reachable by anyone with the address. - Billing: per second. Stopping a pod releases the GPU and keeps the volume disk, which costs $0.20 per GB per month while stopped [21]. For example, a 20 GB volume would be $4.00 a month (arithmetic on that rate). Terminating deletes everything not stored on a network volume [20]. Download first. For a one-off part, download then terminate. After your first terminate, check the console to confirm nothing is still billing.
6. Prompts to give your AI assistant
Paste these into whichever assistant you use. Replace the parts in [brackets]. Every prompt starts with the same rules: ask before spending money, shut the pod down when finished, and run a printability check before recommending a print. Keep those rules.
Prompt 1: set up the toolchain
Rules: ask me before you create, start, resize, stop or delete anything that costs money, and tell me the hourly price first. Do not open a public web server on the pod. When we are finished, tell me which files to download, wait for me to confirm, then ask whether to stop or terminate the pod and say what each choice keeps and costs. Do not claim a part is safe or strong; say what you checked and what you did not. Before you recommend printing anything, run a printability check and tell me what failed. I want to design 3D-printable parts. First ask which printer, nozzle and filament I use. If my own computer can run the tools, tell me how to install them there and do not rent anything. If not, propose the cheapest pod that can run OpenSCAD, build123d, trimesh and OrcaSlicer's command line (a CPU pod if possible) and quote the hourly price before creating it. After I approve, install the tools and print the version of each one, and list any that failed to install.
You should get: a table of installed tools and versions, with failures listed. Check: ask for the output of openscad --version and the slicer's help text, and compare them with the table.
Prompt 2: a bracket, stand or organizer from measurements
Rules: ask me before you create, start, resize, stop or delete anything that costs money, and tell me the hourly price first. Do not open a public web server on the pod. When we are finished, tell me which files to download, wait for me to confirm, then ask whether to stop or terminate the pod and say what each choice keeps and costs. Do not claim a part is safe or strong; say what you checked and what you did not. Before you recommend printing anything, run a printability check and tell me what failed. Design a [wall bracket / phone stand / drawer organizer] as a parametric script. My measurements in millimetres: [list]. Nozzle [0.4] mm. Put every dimension in a named variable at the top, use a minimum wall of [1.6] mm, and avoid overhangs steeper than 45 degrees unless you tell me. Export STL and 3MF, then report the bounding box, the volume and whether the mesh is watertight. Show me images from at least three angles, and list which numbers you assumed.
You should get: a script with named dimensions, an STL and 3MF, a size report and images. Check: compare the bounding box with your measurements, look at the images, and preview the layers in your slicer before printing.
Prompt 3: an enclosure for a board
Rules: ask me before you create, start, resize, stop or delete anything that costs money, and tell me the hourly price first. Do not open a public web server on the pod. When we are finished, tell me which files to download, wait for me to confirm, then ask whether to stop or terminate the pod and say what each choice keeps and costs. Do not claim a part is safe or strong; say what you checked and what you did not. Before you recommend printing anything, run a printability check and tell me what failed. Design a two-part enclosure for [board name]. First ask me for the board's official mechanical drawing or measure it with me; do not guess hole positions. Put mounting posts at the real hole positions, add openings for [ports], vent slots, and a lid held by [screws / snap fit]. Keep every clearance as a named variable. Tell me I must test-fit with a cheap partial print before mounting the board, and list which dimensions came from a source and which you assumed.
You should get: a two-part parametric enclosure and a table of sourced versus assumed dimensions. Check: print only the base or a thin slice first and test-fit the real board with calipers.
Prompt 4: generate a figurine with an open AI model, then check it
Rules: ask me before you create, start, resize, stop or delete anything that costs money, and tell me the hourly price first. Do not open a public web server on the pod. When we are finished, tell me which files to download, wait for me to confirm, then ask whether to stop or terminate the pod and say what each choice keeps and costs. Do not claim a part is safe or strong; say what you checked and what you did not. Before you recommend printing anything, run a printability check and tell me what failed. I want a decorative figurine of [description] about [60] mm tall. Before doing anything, tell me which open 3D generation model you propose (for example TripoSR or TRELLIS), its licence terms including any regional or commercial limits, how much GPU memory it needs, which GPU you would rent and the hourly price. Wait for my yes. Then generate it, scale it to [60] mm tall, and check: watertight yes or no, number of separate bodies, thinnest wall, overhang areas and bounding box. Do not call it printable unless the checks pass, and tell me what failed.
You should get: a scaled mesh and a check report with failures listed. Check: open the mesh yourself and look for floating pieces, thin spines and open holes, and read the model's licence on its page.
Prompt 5: repair and orient a mesh
Rules: ask me before you create, start, resize, stop or delete anything that costs money, and tell me the hourly price first. Do not open a public web server on the pod. When we are finished, tell me which files to download, wait for me to confirm, then ask whether to stop or terminate the pod and say what each choice keeps and costs. Do not claim a part is safe or strong; say what you checked and what you did not. Before you recommend printing anything, run a printability check and tell me what failed. I will give you [file]. Check it with trimesh for watertightness, inverted normals, duplicate vertices and separate bodies. Repair what can be repaired, leave the original untouched, and write a second repaired file. Tell me exactly what changed, with counts. Confirm the size in millimetres did not change. Then recommend a print orientation for FDM, with no supports if possible, explain the trade-off, and export the oriented file.
You should get: an original and a repaired file, a change list and an orientation recommendation. Check: load both in your slicer and look for features that deformed after hole filling.
Prompt 6: estimate print time and filament with a slicer
Rules: ask me before you create, start, resize, stop or delete anything that costs money, and tell me the hourly price first. Do not open a public web server on the pod. When we are finished, tell me which files to download, wait for me to confirm, then ask whether to stop or terminate the pod and say what each choice keeps and costs. Do not claim a part is safe or strong; say what you checked and what you did not. Before you recommend printing anything, run a printability check and tell me what failed. Slice [file] with [OrcaSlicer / PrusaSlicer] on the command line, using the profile for my [printer, nozzle, filament, layer height 0.2 mm]. If you do not have an exact profile, ask me for my exported profile, or say clearly that the numbers are a rough guess. Report estimated print time, filament length and weight, number of layers and any warnings, and show the exact command you ran. If the slicer cannot run on this machine, say so instead of guessing.
You should get: time, filament and warnings, plus the command line. Check: slice the same file with the same settings in your own slicer; the numbers should be close, though real print time can still differ.
Prompt 7: fix a fit problem
Rules: ask me before you create, start, resize, stop or delete anything that costs money, and tell me the hourly price first. Do not open a public web server on the pod. When we are finished, tell me which files to download, wait for me to confirm, then ask whether to stop or terminate the pod and say what each choice keeps and costs. Do not claim a part is safe or strong; say what you checked and what you did not. Before you recommend printing anything, run a printability check and tell me what failed. My printed part [describe] measured [size] but should be [target size]; [a lid is too tight / a screw hole is too small]. My measurements in mm: [list]. Change only the relevant clearance variables, explain what you changed and by how much, regenerate the file, and list which other dimensions are affected. Suggest a small test piece with a range of clearances so I can find the right value for my printer before I reprint the whole part.
You should get: an updated parametric file with a change list and a small clearance test piece. Check: print the test piece, measure it with calipers, and set the clearance to the value that fits.
Prompt 8: wrap up and clean up
Rules: ask me before you create, start, resize, stop or delete anything that costs money, and tell me the hourly price first. Do not open a public web server on the pod. When we are finished, tell me which files to download, wait for me to confirm, then ask whether to stop or terminate the pod and say what each choice keeps and costs. Do not claim a part is safe or strong; say what you checked and what you did not. Before you recommend printing anything, run a printability check and tell me what failed. We are done. List every file you made on the pod with its size and tell me how to download each one to my device. Wait for me. After I confirm I have the files, list my pods with their status and cost per hour, tell me exactly what stopping and terminating would each keep and charge, and ask which I want. Do not stop or terminate anything without my yes.
You should get: a file list, download instructions and a pod status list. Check: open the Pods page in the RunPod console and confirm nothing is running after you finish.
7. Keep it safe
- Load-bearing and safety-critical parts: do not rely on an AI-designed or AI-generated part where a failure could injure someone or damage property, such as climbing, vehicles, medical use, child safety, pressure or heat. No assistant can certify strength, and FDM parts are weaker along layer lines.
- Preview before you print: look at the model, then at the slicer's layer preview. A script or a generated mesh will not always print on the first try, so plan a test print.
- Electrical enclosures: a printed plastic box is not a certified electrical enclosure. For mains voltage, heat sources or batteries, check the material's datasheet and use a properly rated enclosure.
- Food contact: Prusa says food-safe FDM printing is not as simple as it might seem and does not recommend using FDM prints as food containers. Layer lines can harbour bacteria, and colourants and additives may not be inert [23].
- Who owns AI-generated shapes: this is legally unsettled and depends on the country. The US Copyright Office has published a report on the copyrightability of generative AI outputs [24], but we did not read its conclusions. Check the model licence, and take legal advice if you plan to sell prints. This is not legal advice.
- Public addresses: do not put private designs on an open pod port. See section 5.
- Always terminate a finished pod and check the pod list.
8. What we have not tested yet
This guide is built from documentation and licence pages, not from a finished print of our own. Before we call anything proven, these still need a real hands-on test:
- The whole code-CAD route on a CPU pod: installing OpenSCAD, build123d, trimesh and a slicer on the Ubuntu templates, how long it takes, and which versions you get.
- Whether OrcaSlicer or PrusaSlicer run on a pod with no display, and how to get a matching printer profile without the desktop app.
- Preview images from a pod without a display, from OpenSCAD or Blender.
- The iPad route end to end: opening the web terminal and JupyterLab, and downloading an STL or 3MF in Safari.
- TRELLIS.2 on a 24 GB card, and Hunyuan3D-2.1 with shape plus texture, including the low-memory flag. Install problems and the cost per model.
- How often AI-generated meshes pass a slicer check without repair. We plan to generate a set of models with each tool and report the share that needed repair.
- Which repair tools fix which defects, and whether they keep the scale.
- How close slicer estimates are to real print time on an actual printer.
- Whether an assistant can carry clearance changes through an enclosure redesign, checked with a physical print.
- The exact charges after a stop and after a terminate, read from the console for a test pod.
When we run these, we will publish the measurements with the date, the hardware and the bill.
Want to see them when they land? Follow @3DeltaD on X. That is where new guides and updates are posted.
Next: get it made
First check whether someone has already designed it. Then print it at home or send it to a print service. We earn nothing from these links, and we have not ordered from them ourselves. Get prices from each service's own quote page. Sharing a model made with an AI generator? Each library has its own rule: read it before you upload, and label the model honestly.
- Printables and MakerWorld: large free model libraries. Search before you design. Reviewers mention being asked to sign in to download from some libraries.
- Craftcloud: upload once and compare prices from many print services, including nylon (SLS, MJF), resin and metal. Its site says no registration is needed for a quote.
- JLC3DP or PCBWay: order nylon, resin or metal prints directly. Allow for shipping and import charges. Some public reviews of their sister board services mention refund and shipping problems, so start small.
- Xometry: for engineering parts where material and accuracy matter. Built for engineers, so it is not priced for hobby parts.
What to have ready
- An STL or 3MF file for a home printer or a model library.
- A STEP file as well if you designed the part with code CAD; print services prefer it for exact shapes.
- The real size in millimetres. An STL file does not store units, so check the size shown in the upload preview.
- One solid, closed ("watertight") shape, with no loose floating pieces.
- What the part is for, so you can pick a material: indoors, outdoors, near heat, flexible or stiff.
Before you order
- The size in the service's preview matches your measurements.
- The thinnest wall is thicker than the minimum that service lists for that material. Look up their design guide; your assistant does not know it.
- If you can, print a cheap plastic test at home or ask a friend before paying for nylon or metal.
- It is not a part where a failure could hurt someone: nothing load-bearing, no child-safety, vehicle, climbing, pressure or mains-electrical use. No food contact.
- For an AI-generated shape: you looked at it from every side, read the generator's licence, and are not copying a branded character.
- You read the full total including shipping, and know import charges may come later.
- You placed the order yourself.
Print this checklist on one page, or tick it off on screen.
Want a pod for AI mesh generation?
Make a RunPod account and add a small balance such as $10, or paste this into your AI assistant: Read agents.runpod.io and set up Runpod for me. Then use Prompt 4 above.
Referral link: we may earn credits if you sign up.
- RunPod's sign-up page opens in a new tab. You make the account yourself, on RunPod's site. The link carries our referral code.
- To rent anything you first add a balance. That is a payment to RunPod, so start small, such as $10. RunPod's referral terms describe a bonus credit for a new account that signs up with Google and loads at least $10. Read the terms for the amounts and regions.
- Nothing starts on its own. You choose a machine, or your assistant proposes one and, as the starter messages tell it to, waits for your "yes, go". Check the price per hour before anything is deployed.
- A running machine bills by the second. When you finish, download your files, terminate the pod (a stopped pod still bills for its disk) and check the console shows nothing running.
Not ready? That is fine. Most of this guide needs no pod: start with the prompts, or follow @3DeltaD on X and come back later.
Sources
- OpenSCAD command line (Wikibooks): en.wikibooks.org. OpenSCAD licence: github.com/openscad/openscad (COPYING)
- CadQuery: github.com/CadQuery/cadquery; pypi.org/pypi/cadquery
- build123d: github.com/gumyr/build123d; export docs: build123d.readthedocs.io
- Headless FreeCAD (mirror of the FreeCAD wiki article): reqrefusion.github.io. FreeCAD: github.com/FreeCAD/FreeCAD
- Blender bpy on PyPI: pypi.org/pypi/bpy. Blender command-line arguments: docs.blender.org
- trimesh: pypi.org/pypi/trimesh; repair docs: trimesh.org/trimesh.repair.html
- PyMeshLab: github.com/cnr-isti-vclab/PyMeshLab
- OrcaSlicer, CLI mode: orcaslicer.com/wiki/cli/cli_mode; CLI actions: orcaslicer.com/wiki/cli/cli_actions
- PrusaSlicer man page (Debian package): manpages.debian.org
- TRELLIS model page: huggingface.co/microsoft/TRELLIS-image-large. README, as read from a mirror of the project: github.com/VAST-AI-Research/TRELLIS
- TRELLIS.2: github.com/microsoft/TRELLIS.2; model page: huggingface.co/microsoft/TRELLIS.2-4B
- Hunyuan3D-2.1: github.com/Tencent-Hunyuan/Hunyuan3D-2.1; licence: LICENSE; model page: huggingface.co/tencent/Hunyuan3D-2.1
- TripoSR: github.com/VAST-AI-Research/TripoSR
- Stable Fast 3D model page: huggingface.co/stabilityai/stable-fast-3d; code: github.com/Stability-AI/stable-fast-3d
- RunPod docs, Pods overview: docs.runpod.io/pods/overview. Template listing read through RunPod's own tools within the last 30 days; PyTorch image tag: hub.docker.com (runpod/pytorch)
- RunPod docs, Connection options: docs.runpod.io/pods/connect-to-a-pod. Uploading data (RunPod blog): runpod.io/blog/transfer-data-into-runpod
- RunPod docs, Expose ports: docs.runpod.io/pods/configuration/expose-ports
- RunPod docs, Transfer files: docs.runpod.io/pods/storage/transfer-files; runpodctl send: docs.runpod.io/runpodctl/reference/runpodctl-send
- RunPod docs, Use SSH: docs.runpod.io/pods/configuration/use-ssh
- RunPod docs, Manage pods (stop and terminate): docs.runpod.io/pods/manage-pods
- RunPod docs, Pod pricing and storage: docs.runpod.io/pods/pricing
- RunPod GPU selection guide (24 GB and 48 GB card tiers): github.com/runpod/runpod-plugins-official
- Prusa Knowledge Base, Food safe FDM printing: help.prusa3d.com
- US Copyright Office, Copyright and Artificial Intelligence: copyright.gov/ai
Licence and rate information was read within the last 30 days and can change; check each page before you rely on it. GPU rates are RunPod on-demand rates from its API. Product names belong to their owners and are used only to describe what they are.
