One Photo In, a 3D Object Out: What Local Image-to-3D Actually Gives You
Idol turns one photo into a 3D object and topped Hugging Face trending. It also writes a file no slicer accepts. What each image-to-3D tool really needs.
Something plastic has broken — the knob off a washing machine drawer, the clip that holds a vacuum hose to its handle — and the part is not sold on its own. You have a phone, a printer and a photograph of the thing. So when a local image to 3D model turns up that takes one picture and hands back a 3D object, and its downloads pass 14,613 in a month, it looks like the gap just closed.
It did not close, and the reason is not quality. It is the file format, and you can know whether any of these tools will work for you before you install a single one of them.
What a local image to 3D model actually hands you
Two different things travel under the same description, and almost everything below depends on telling them apart.
A mesh is a surface: a shell of triangles with an inside and an outside. It is what a slicer expects, what a game engine expects, and what you can scale, hollow, cut in half and print. The common formats are OBJ, STL and 3MF.
A Gaussian splat is not a surface. It is a cloud of soft coloured blobs, each with a position, a size and an orientation, which a renderer draws front-to-back so that the result looks like a solid object from any angle you orbit to. It is a genuinely clever way to store the appearance of a thing. It stores no shell at all. Point a slicer at one and there is nothing for it to slice, because there is no surface.
Both of them routinely arrive in a file ending .ply, which is where the confusion starts.
Trying Idol without installing anything
The model that caused this week’s spike is Idol, published to Hugging Face on 28 October by Speridlabs. There is a hosted demo, and it is worth starting there — partly because it takes two minutes, and mostly because it is the only route open to you if you are on a Mac, on an AMD card, or on a laptop with no discrete GPU at all. The local install needs an NVIDIA GPU with CUDA, PyTorch 2.5 or later and Python 3.10 or later, and that requirement is not negotiable.
Open the Space, and the layout is one screen: an input image on the left, an empty viewer on the right.
- Pick an example or upload your own photo. The gallery along the bottom has a loaf of bread, a castle, a bar stool, a gingerbread house.
- Set the input type. Synthetic for a render or a generated image, real for a photograph. The only thing this changes is how much room the crop leaves around the object — synthetic fills the frame more tightly, because a render usually arrives already isolated.
- Press Generate, and wait.
The result is genuinely impressive. Running the demo’s own backpack example, the object comes back orbitable, and the angles the photograph never showed hold their shape and their colour rather than smearing. Inferring the back of a thing from its front is the hard part, and it does that well.
The demo’s own progress counter reported 26.5 seconds for that run, on the project’s free hosted GPU, on 5 November. Hold onto that number.
The two-second claim and the table underneath it
Idol’s model card opens by saying it “runs on consumer GPUs, using 7.2 GB of VRAM and can run in under 2 seconds.”
Further down the same page is a table of measured request times, at 50 steps and 262,144 Gaussians, taken as the median of eight images with identical code and settings on every card. Ten GPUs are listed. Here they are in full:
| GPU | VRAM | Request time |
|---|---|---|
| B300 | 269 GB | 3.4 s |
| B200 | 179 GB | 3.6 s |
| H200 | 140 GB | 6.6 s |
| H100 | 80 GB | 6.7 s |
| RTX 5090 | 32 GB | 14.1 s |
| A100 | 80 GB | 16.3 s |
| L40S | 45 GB | 16.6 s |
| RTX 4090 | 24 GB | 19.1 s |
| RTX 3080 | 10 GB | 46.2 s |
| L4 | 22 GB | 54.6 s |
Nothing on that list reaches two seconds. The fastest card in it, a B300, is a datacentre accelerator with 269 GB of memory, and it takes 3.4. The quickest consumer card is an RTX 5090 at 14.1 seconds. A 4090 takes 19.1. An RTX 3080 — which does clear the 7.2 GB requirement, with 10 GB — takes 46.2 seconds, or roughly twenty-three times the headline.
The page does not say which setting produces the sub-two-second figure, so there is no point guessing at one. The useful reading is narrower and more practical: runs on consumer GPUs and under 2 seconds are both true sentences, and they are never true at the same time. If you are budgeting your evening around this, budget in the 15-to-50 second range per object, plus a slower first run while the model compiles.
None of this makes Idol a bad model. On the same H100 it finishes in 6.8 seconds against TRELLIS’s 7.1, and beats TRELLIS on every one of the seven quality scores its benchmark reports. Across all four models in that table it takes six of the seven, losing only PSNR to LGM — which the page itself flags, noting LGM wins that metric “despite visibly blurry results”. The complaint is with one sentence at the top of a page, which is a different thing from a complaint about the work.
Why you cannot print what Idol makes
Here is the part the ranking pages do not mention.
Idol writes exactly one 3D file per image: a standard 3DGS .ply, which its own documentation describes as something to “open in any Gaussian splat viewer”. Alongside it you get the cropped input image and a JSON file holding the predicted camera position. That is the complete output. There is no mesh, and there is no option for one.
PrusaSlicer imports 3MF, STL, OBJ, STEP and SVG. The previous version’s list is STL, STEP, 3MF, OBJ and AMF. PLY appears on neither.
So this is not a quality problem that a bit of cleanup in Blender will fix. It is not a mesh with holes in it, needing repair. There is no mesh. Projects exist that try to extract a surface from a splat afterwards, but that is a separate job with its own tools and its own failure modes, and sending somebody who wanted a replacement knob down an untested conversion chain is how a person loses an evening.
If what you want is something a printer can take, the straightforward move is to use a tool that produces a mesh in the first place.
The three alternatives, and what each one costs you
Every figure here comes from the project’s own README or its source code, read on 5 November, rather than from the comparison pages, which largely quote each other.
TripoSR is the low-requirement option. Its README states plainly that the default options take about 6 GB of VRAM for a single image. Reading run.py settles the output question: it writes mesh.obj by default, with GLB as the alternative. OBJ goes straight into a slicer. It is MIT licensed, it has 7,289 stars, and it was last pushed on 4 June 2026 — five months quiet, which is slow but not abandoned. The meshes are rough. For a functional bracket or a clip, rough is often fine.
TRELLIS, from Microsoft Research, is the one to reach for if you have the hardware. It outputs radiance fields, 3D Gaussians and meshes, which makes it the only tool here that lets you look at the splat and print the shell from the same run. The cost is in the first line of its requirements: an NVIDIA GPU with at least 16 GB of memory is necessary. That rules out the 10 GB RTX 3080 from the table above, an 8 GB 4060 Ti, and most laptop cards. It is verified on A100 and A6000 hardware, Linux is the only tested operating system, and Windows support is described as only partly tested. It is MIT for the models and most of the code, with some submodules carrying their own terms. Last pushed 27 October 2026, so it is alive.
Hunyuan3D-2 is the one the comparison pages call the quality pick. It has 15,332 stars and was last pushed on 28 October 2025 — over a year ago. Its licence is the reason it gets its own section below, and the reason this piece makes no recommendation about its output.
Idol is 1.8B parameters across six components, peaks at 7.2 GB, is brand new, produces the best-looking result of the four in its own tests, and gives you a splat. Which it is honest about: “A single image to a 3D Gaussian splat” is the opening line of its own page. It is just not the line anyone quotes.
The licences are not what the badges say
Idol’s Hugging Face page displays apache-2.0, and the repository’s own weights genuinely are. But the model will not start until you have been granted access to two gated third-party repositories, and the terms on those are not Apache.
The first is Meta’s DINOv3 ViT-H+/16 image encoder, listed under a dinov3-license — Meta’s own terms, not an OSI licence — and gated behind agreeing to share your Hugging Face account information with Meta. The second is FLUX.2-dev, from which Idol downloads only the VAE. That page states that the model “falls under the FLUX Non-Commercial License”, with separate self-hosted commercial terms offered elsewhere.
So the practical position for anyone planning to sell what comes out is: the badge says Apache 2.0, and the dependency chain says read two more agreements first. This site has run into the same shape before, with a 1.7B model whose open licence carried a revenue cap nobody reads. The badge is a summary of one repository, not of the thing you are about to run.
Hunyuan3D goes considerably further. Its licence opens, in capitals, on line three:
THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA AND IS EXPRESSLY LIMITED TO THE TERRITORY, AS DEFINED BELOW.
Section 1(l) defines that Territory as the world minus the EU, the UK and South Korea. Section 2 grants its rights “for the Territory only”. Section 5(c) states that use outside the Territory “is unlicensed and unauthorized under this Agreement” — and it applies that to the output and results, not only to the weights. There is also a commercial threshold at a million monthly active users, which will not trouble most readers, but the territorial exclusion very much will.
A reader in London or Berlin following a listicle that calls Hunyuan3D the best open-source choice is being pointed at something they have no licence to run at all. That is not a technicality buried in clause 47. It is the third line of the file, in capitals, and it is the single best argument for the habit of opening a project’s own LICENSE before its README.
So which one should you actually use
The decision is short once the formats are straight.
You want something printable and have a 6 GB NVIDIA card: TripoSR. OBJ out, MIT licence, no gated dependencies, rough results that a slicer will accept today.
You want the best result and have 16 GB and Linux: TRELLIS. Mesh and splat from the same run, MIT, and still being maintained.
You are on a Mac, on an AMD card, or on a laptop with integrated graphics: the hosted Space, and go in knowing that what comes back is a splat to look at and orbit, not a part to print. If local AI on a Mac is the actual question, how local file-search tools for the Mac held up when their claims were checked and offline OCR using what macOS already ships are better places to spend an afternoon, because those run.
You are in the EU, the UK or South Korea: Hunyuan3D is not on your list, whatever the comparison pages say.
What stands out after a day in these projects’ own documentation is how little of the decision turns on model quality. All four produce something recognisable from a single photograph. The questions that actually decide the outcome are whether your card has the memory, whether a mesh comes out of the other end, and whether the licence covers the country you live in — and all three are answerable in ten minutes, before any download starts. That holds well beyond 3D: the same ten minutes is what separates a portable local LLM setup that works from one that does not.
Check the output format first. The model is the easy part.
Sources
Every figure in this article traces to one of these. Dates are when the page was read.
- speridlabs/idol model cardHugging FaceAccessed Nov 5, 2026Official docs
- Idol — the official Hugging Face SpaceSperidlabsAccessed Nov 5, 2026Official docs
- FLUX.2-dev model cardBlack Forest LabsAccessed Nov 5, 2026Official docs
- DINOv3 ViT-H+/16 model cardMetaAccessed Nov 5, 2026Official docs
- TripoSRVAST AI ResearchAccessed Nov 5, 2026Official docs
- TRELLISMicrosoft ResearchAccessed Nov 5, 2026Official docs
- Tencent Hunyuan 3D 2.0 Community License AgreementTencentAccessed Nov 5, 2026Official docs
- Supported file formatsPrusa ResearchAccessed Nov 5, 2026Official docs
Written by
Full-stack developer who builds and runs AI automation systems in production. Runs local LLMs on personal hardware, builds N8N pipelines that actually ship, and deploys on Cloudflare Pages. Every guide on Pipeline Monk is tested on real consumer hardware — a Ryzen 7 5800HS with 16GB RAM and a GTX 1650. If it works on that, it works.