Idol Turns One Picture Into a 3D Object for Free. The Back Side Is a Guess, So Shoot the Front for It
Show a new 3D model a photo of a teapot and it will hand you the whole teapot: spout, handle, and the back you never photographed. That back is not reconstructed. It is invented, a confident guess about what teapots usually look like behind.
That one fact decides how useful Idol is to you, more than any speed figure or benchmark table. Idol landed on Hugging Face on October 28 from a small lab called Speridlabs. It turns a single image into an object you can spin around, it runs on a mid-range graphics card, and its licence lets you use what you make in paid work. Within two days the community had it running inside ComfyUI and on Apple Silicon Macs.
So the interesting question is not "does it work." It mostly does. The question is what kind of picture gives it the right things to guess from, and what you can actually do with the object once you have it.
What Idol makes, in plain terms
Most 3D tools hand you a mesh: a surface built from flat triangles, the kind you can texture, light and rig in Blender or a game engine. Idol does not. It outputs a Gaussian splat, which is a cloud of up to about a million soft, coloured, semi-transparent blobs. Viewed from any angle, those blobs blend into what looks like a solid object.
Splats are very good at looking real. Fuzzy edges, fabric, foliage and glossy surfaces come through better than they do on a quick mesh. They are less good at being worked on. You cannot easily unwrap one for texturing or attach it to a skeleton.
That gives you the first practical rule. Idol is a tool for seeing an object from every side, fast. It is not yet a tool for producing a game-ready asset.
You hand it a photo, a render or a generated image. It removes the background, crops to the object, and writes out a standard 3D splat file (a .ply). You choose how dense the object is, from about 33,000 blobs for a quick look to just over a million for detail.
Why it matters that this one is open
Image-to-3D tools are not new. What is new here is the combination of three things in one free download.
It runs on ordinary hardware. Speridlabs lists a peak of 7.2 GB of graphics memory on its model page. The community ComfyUI version reports about 6 GB, tested on an RTX 3090 and an RTX 4090. That puts it within reach of a lot of gaming cards already sitting in artists' machines.
The licence lets you sell the work. The model is released under Apache 2.0, a permissive licence that allows commercial use. One caveat: before first run you have to accept the terms on two other Hugging Face models Idol depends on (Meta's DINOv3 image reader, and an image decoder from FLUX.2-dev). Each has its own terms. Read them once, especially if a client will see the results.
It does not take long. Speridlabs' own timing table runs from 3.4 seconds on a data-centre card to 54.6 seconds on an L4, a much more modest card, for a medium-detail object. The model page also says "under 2 seconds" in its headline, which only applies at the very top of that table. Plan for something closer to a minute on whatever you own, and enjoy it if it comes back faster.
On the Mac side, a community build for a local app called mlx-serve reports about 40 seconds per object on an M-series Ultra chip, using about 3.6 GB of memory. That is the porter's own measurement on one machine.
The guess is the whole game
Here is where the teapot comes back.
Idol sees one view. Everything else, it fills in from patterns it learned. The ComfyUI port's author puts it bluntly in the README: the back is invented from the front, so what the picture shows decides it. The same note says a three-quarter view tells Idol more than a straight-on front shot.
Think about why. A dead-centre front photo of a mug hides its depth and the handle's shape. A three-quarter view shows two faces, the curve of the body, and where the handle meets it. You have handed the model evidence instead of asking it to imagine.
This is the angle launch coverage tends to skip. The speed and the benchmarks are about the model. The quality of your result is mostly about your input. A product photographer, a prop maker or a concept artist already knows how to choose the angle that explains an object. That skill now feeds straight into a 3D result.
Speridlabs' benchmark is the company's own, run against three other open tools. Idol scores best on the views it had to invent, which is exactly what matters here. It is still a guess. Unusual objects, hidden mechanisms and anything with text on the back will come out as plausible fiction.
Put this into practice
The quickest route costs nothing and needs no install.
1. Try the free demo first. The Idol Space on Hugging Face takes an upload, lets you set detail and the random starting number (the seed), and shows the result in a viewer you can spin and download from. It runs on shared hardware, so you may wait in a queue.
2. Shoot or pick the right picture. One object, clearly separated from the background, seen from about three-quarters. Even, soft light helps, because whatever shadows and highlights are in the photo get baked into the object's colours.
3. Run it locally in ComfyUI if you like the result. The community ComfyUI-Idol pack needs ComfyUI 0.43 or newer. Install "Idol (image to 3D splat)" from ComfyUI-Manager, download the model files listed on the Idol-ComfyUI page (about 2.6 GB at full precision, half that for the smaller version, plus helper files from VAST-AI's TripoSplat repo), then open the idol_image_to_splat template.
4. Set the mode to match your picture. Use synthetic for renders and AI-generated images. For real photographs, set mode to real on both Idol nodes. The README says the photo setting crops more loosely around the object.
5. Check what the model sees. The template includes a preview of the cleaned-up input. The object should sit alone on black. If background is left over, it will turn into stray blobs in 3D.
6. Try a few seeds. If the back looks wrong, run it again with a different seed before you change anything else. The README suggests exactly this.
You get the .ply file in ComfyUI's output/3d folder, plus a render from the original viewpoint and a turntable video, which on its own is a handy client preview.
What to do with the object afterwards
This is where splats still ask for patience.
For cleanup and sharing, SuperSplat is a free, open-source splat editor from PlayCanvas that runs in a browser with nothing to install. It is a good place to inspect the result and tidy it before you send it anywhere.
For Blender users, the news is close but not quite here. Blender 5.3, which Digital Production reports is due on November 10 and is in alpha now, imports splat files natively and can render them in Workbench, EEVEE and Cycles. The same report notes it cannot export splats yet, colour handling is not fully correct, and performance still needs work. Until then, third-party add-ons fill the gap.
If you need a mesh you can texture and rig, Idol is the wrong finishing tool. Use it to decide the shape, then model or retopologise properly.
Where it breaks
The hidden side is fiction. Fine for a mood board, a blockout or a turnaround reference. Risky for a product listing, where a customer might believe the back you never photographed.
Lighting is baked in. Shadows and reflections in your photo become part of the object's colour. Relight it in a new scene and those old shadows come along.
Official code wants an NVIDIA card. Speridlabs' own tools need an NVIDIA card with CUDA, NVIDIA's computing software. Mac users depend on community builds, which are one person's work tested on one machine.
The ports are young. The ComfyUI pack was tested on two cards. Its author also notes that the same seed gives a different object in ComfyUI than in Speridlabs' own tools, so do not expect to reproduce a demo result exactly.
Speed claims need reading closely. The two-second headline belongs to a data-centre card. Budget your time around the slow end of the table.
Splats are not the format most pipelines expect. Game engines and many renderers need a plugin or a conversion step before they will touch one.
A new first step for 3D
The useful way to think about Idol is as a sketchbook for objects. A concept artist can turn a painted prop into something the art director can spin around. A product designer can check proportions from every side before a single CAD hour. An indie developer can block out a room full of placeholder props in an afternoon, then replace the ones that survive.
None of that needs the hidden side to be true. It only needs it to be plausible enough to make a decision.
So pick one object on your desk this week, photograph it from three-quarters in soft window light, and put it through the free demo. Then photograph it straight on and run it again. The difference between the two backs will tell you more about how to use this tool than any benchmark will.