ComfyUI can run Hunyuan3D, TRELLIS and the lighter 3D models behind one node graph, but the weights, the VRAM budget and the trip from a voxel grid to a printable mesh are three separate problems.
ComfyUI did not start as a 3D tool. It became one in stages: community node packs wrapped research repositories first, then core nodes for Hunyuan3D 2.0 shipped with the application, and later a built-in 3D viewer and a GLB export node. Today the same graph that renders a concept image can hand it to a shape model and write a mesh file without leaving the browser tab.
The graph does not hide the plumbing. A working setup means picking a model family your card can hold, downloading tens of gigabytes of weights into the folders each loader expects, and accepting a dense triangle mesh instead of a finished asset.
ComfyUI is a node graph runner with a queue, a model cache and a browser UI, and its 3D support is much thinner than its image support. What you get is reproducible execution: an image, a shape model, a decoder and an export node chained into a workflow that saves as JSON. What you do not get is topology, UV quality, rigging or print preparation.
| Family | Input | Typical VRAM | Output |
|---|---|---|---|
| Hunyuan3D 2.0 / 2.1 full | 1 image or up to 4 views | 16 GB; 24 GB with PBR texturing | Textured mesh, good silhouette, dense triangles |
| Hunyuan3D mini and turbo | 1 image | 8-12 GB with offloading | Fewer steps, softer detail |
| TRELLIS and TRELLIS 2 | 1-4 images | 16 GB and up | Sparse-latent mesh, UVs after painting |
| Stable Fast 3D | 1 image | around 6-8 GB | UV-unwrapped mesh with a baked texture |
| TripoSR | 1 image | around 6 GB | Fast untextured mesh, vertex colours |
Choosing between them is mostly a VRAM decision. For a printable silhouette, Hunyuan3D is the best documented path and its mini and turbo checkpoints stay inside 12 GB. For UVs and PBR maps without a separate baking pass, TRELLIS and Stable Fast 3D deliver, at the cost of memory or sharper geometry. TripoSR is the small-card fallback: seconds per run, vertex colours only.
In Kijai's wrapper the graph stays short: Hy3DModelLoader loads the DiT checkpoint and outputs a pipeline plus a VAE, Hy3DGenerateMesh samples the latents, and Hy3DVAEDecode decodes them into a mesh at an octree resolution you choose. Turbo checkpoints typically run 5-8 steps, the full model around 50. The core workflow gets there with EmptyLatentHunyuan3Dv2, Hunyuan3Dv2Conditioning and VAEDecodeHunyuan3D, then VoxelToMeshBasic, whose density threshold places the surface. Texturing is a separate, heavier stage.
TRELLIS predicts a structured 3D latent rather than a voxel grid, and its ComfyUI packs accept one to four conditioning images, which shortens the multiview path. The price is memory and setup: those packs follow a fast-moving research repository, compiled wheels such as spconv must match your torch build, and a comfortable run starts around 16 GB. Stable Fast 3D trades the other way: one forward pass, roughly 6-8 GB, texture already baked onto UVs.
Two settings do most of the visible work. Octree resolution decides how much surface detail survives the voxel stage, and the density threshold in VoxelToMeshBasic decides where the surface sits: lower values inflate the model, higher values can disconnect thin parts. No sampler setting rescues a photo with two overlapping subjects.
| Card memory | Runs without tricks | What you give up |
|---|---|---|
| 8 GB | TripoSR, Stable Fast 3D, Hunyuan3D mini shape-only | Quantised or offloaded runs, minutes per mesh, no paint stage |
| 12 GB | Hunyuan3D turbo geometry, small TRELLIS runs with offload | Longer sampling, lower octree resolution, occasional stalls |
| 16 GB | Full Hunyuan3D 2.0 shape model, TRELLIS image-to-3D | The PBR texture stage, competing for the same memory |
| 24 GB | Hunyuan3D 2.1 with the paint model, TRELLIS 2, batches | Nothing structural; wall-clock time is the limit |
Caching behaviour matters as much as capacity, because ComfyUI keeps models resident between runs and does not always release them. Starting with --lowvram, --novram or --disable-smart-memory changes what stays in memory, and quantised community conversions of the DiT checkpoints trade surface quality for a tier of VRAM. Offloading to system RAM works: the job finishes, it just spends its time copying tensors over PCIe.
The 3D models wrapped for ComfyUI are image-conditioned, so text-to-3D is a chain: the prompt goes into an SDXL or Flux checkpoint, the render becomes the conditioning image, and the shape model reconstructs it. You iterate on the concept with cheap 2D generations and repair a weak back view by passing one more image as an extra view. Sketch-to-3D follows the same route through a scribble or lineart ControlNet at low denoise strength; a clean silhouette can go straight in, but flat slabs are more common.
| Step | Tool | Typical numbers |
|---|---|---|
| Set the real size | Blender dimensions or the slicer scale field | Generated mesh is about 1 unit tall; set millimetres first |
| Reduce polycount | Blender Decimate, Meshmixer | 200k-1M triangles in, 10-30% ratio out |
| Repair the mesh | 3D-Print Toolbox, Meshmixer, PrusaSlicer, Bambu Studio | No non-manifold edges before slicing |
| Thicken and orient | Blender Solidify, slicer wall settings | 0.8-1.2 mm minimum walls with a 0.4 mm nozzle |
| Texture, if it matters | AI texturing or hand painting | Machine-made UVs suit props, rarely hero assets |
Open3DAI sits at the other end of this trade-off. It runs one local procedural engine plus optional third-party engine adapters you connect with your own API keys, so the same prompt or image can go through the available engines and be compared side by side before export. It writes GLB, glTF, OBJ, STL, PLY and 3MF, and adds 48 format converters, 12 browser tools and a REST API. It does not train or host the third-party models, it does no rigging, animation, retopology, UV unwrapping or photogrammetry, and it does not promise print-ready output without cleanup. Need a reproducible graph with research models wired in: ComfyUI. Need a shape today and a file for the slicer: a hosted generator is usually faster. Engine availability, credit pricing and third-party terms change, so verify them before committing a budget.
Not with the 3D checkpoints themselves. The shape models wrapped for ComfyUI are image-conditioned, so the usual graph puts SDXL or Flux in front of them: the prompt produces a render, and the render becomes the conditioning image. The advantage is that you can iterate cheaply in 2D and only run the 3D model on the frames you like.
The turbo and mini checkpoints are usually runnable in 8-12 GB with offloading, and 16 GB is a comfortable fit for the full shape model. The texture and paint stage is the heavy part and generally wants a 24 GB card, which is why many people generate and texture on different days, or export the untextured mesh and finish it elsewhere.
Because it is one. These models decode a voxel or sparse latent field into a surface, which produces an even, dense triangle mesh with no edge flow and no quad structure. That is fine for printing and for static props, and it is the reason decimation and repair happen in Blender before anything else. No setting in ComfyUI turns that output into animation-ready topology.
Rarely without a pass through Blender or a slicer. Fix the unit scale first, since generated meshes arrive normalised rather than in millimetres, then decimate, repair non-manifold edges and check that thin features are above your printer's minimum wall thickness. PrusaSlicer and Bambu Studio repair some damage on import, but they cannot thicken a finger that is 0.4 mm across.
It depends on which failure costs you more. Hunyuan3D has the shorter install, the more stable node packs and predictable memory use, while TRELLIS handles multiview conditioning more directly and tends to give cleaner surfaces with usable UVs. If you are unsure, install Hunyuan3D first, get one complete run through to an STL, and only then add a second family to compare against.
Send the same prompt or image to several engines, compare the meshes side by side, then export GLB, glTF, OBJ, STL, PLY or 3MF. Creating an account grants free credits and needs no card.
Hunyuan3D 2.x is a shape diffusion model plus a separate texturing stage, and ComfyUI can drive both. Here is how the graph is wired, what each route costs in VRAM, and what the usual error messages mean.
TRELLIS 2 is a 4B-parameter image-to-3D model, and its ComfyUI wrapper depends on compiled CUDA extensions that must match your Python, torch and CUDA versions. This hub covers what it produces, what a machine needs, and how both workflow builds are assembled.
A head-to-head of the three open 3D generators people actually run locally: architecture, mesh fidelity, VRAM and time budgets, licence terms and ComfyUI node support, plus a decision framework you can apply to your own images.