GraphicCardPrices

What GPU do you need for Idol? Turn one image into a 3D splat locally

Illustration of a cartoon knight 3D model on a monitor, with the same knight sketched in a notebook on the desk
Illustration

By GraphicCardPrices. Requirements checked against the official Idol documentation on .

You have one picture of an armchair. Maybe it is a render, maybe a photo from your phone. You want to spin it around in 3D and look at the back, the side the picture never showed.

Idol, an open model from speridlabs, does that on your own PC. One image goes in. A 3D Gaussian splat comes out.

The memory bill is the surprise. The model card puts the peak at 7.2 GB, which is inside the reach of an 8GB graphics card. The catch is in what you get back, and we will get to it.

One picture in, up to a million Gaussians out

Idol does three things to your image. It removes the background, frames the object, and then generates the 3D version, including the sides the camera never saw.

The result is a Gaussian splat: a cloud of soft colored blobs that together look like the object from any angle. The default is 262,144 of them. You can ask for as few as 32,768 or as many as 1,048,576.

Idol writes the splat as a standard 3DGS .ply file, which the model card says opens in any Gaussian splat viewer. There are two input modes: synthetic for renders and generated images, and real for photographs.

Where the 7.2 GB comes from

The model card adds up every part of the pipeline: the denoiser, two decoders, a background remover and two image encoders. Together that is about 1.80 billion parameters and roughly 3.7 to 4.4 GB of weights.

Weights are not the whole story, because the model needs working room while it runs. The card gives the full-request peak in one line:

"It runs on consumer GPUs, using 7.2 GB of VRAM and can run in under 2 seconds."

So an 8GB card clears the published peak, with less than 1 GB to spare. Close the other apps that use your graphics card before you start.

One flag changes the sum. Running the DINOv3 encoder in full precision (--dinov3_dtype fp32) costs "1.7 GB more memory", and the card does not give a new total for that mode. Leave it at the default on a small card.

08GB12GB16GB24GB32GBMinimum 7.2 GBComfortable 10 GBBest 24 GB8GB: RTX 4060, RTX 5060meets Minimum12GB: RTX 3060, RTX 5070meets Comfortable16GB: RTX 5060 Ti, RTX 5070 Timeets Comfortable24GB: RTX 3090, RTX 4090meets Best32GB: RTX 5090meets Best
Will my card run it? Bars show common cards' VRAM; vertical lines mark the official Idol tiers from the table below.
VRAM figures from official documentation. Example cards link to their price pages.
TierVRAMWhat it runsExample cards
Minimum7.2 GBThe full-request peak in the model card: about 3.7 to 4.4 GB of weights and a 7.2 GB peak at the default settings. sourceRTX 4060, RTX 5060, RTX 3060 (12GB version)
Comfortable10 GBThe smallest card in the authors' timing table: the RTX 3080 (10 GB) at 46.2 seconds per request. sourceRTX 3080, RTX 5070, RTX 4070 SUPER
Best24 GBThe RTX 4090 (24 GB) is timed at 19.1 seconds per request and the RTX 5090 (32 GB) at 14.1 seconds. sourceRTX 4090, RTX 5090

What the authors timed, card by card

Most model cards stop at a memory number. This one also lists request times on ten GPUs, measured by the authors with the same code and settings: 50 steps, 262,144 Gaussians, the median over 8 images.

Three of those are cards you can put in a desktop. The RTX 3080 (10 GB) took 46.2 seconds. The RTX 4090 (24 GB) took 19.1 seconds. The RTX 5090 (32 GB) took 14.1 seconds.

Those are the authors' numbers, not ours. They still tell you something useful: every listed card finishes a request in under a minute, so extra money buys shorter waits, not a different result.

Your first splat, step by step

Setup is short, but there is one step people miss: two of the models Idol depends on are gated.

  1. Accept the terms of the two gated models on Hugging Face: facebook/dinov3-vith16plus-pretrain-lvd1689m and black-forest-labs/FLUX.2-dev (only its VAE is downloaded). Then log in once with hf auth login.
  2. Clone the idol-sdk repository then run pip install -e . inside that folder (the trailing dot is part of the command). The model card lists Python 3.10+, PyTorch 2.5+ and an NVIDIA GPU with CUDA.
  3. Run python -m idol.sample --image photo.jpg --mode real for a photograph, or --mode synthetic for a render or generated image. On the first run, Idol's own weights (about 3 GB) and the two encoders download automatically.
  4. Open the .ply file in a Gaussian splat viewer. Idol also saves the cropped input image and a camera file next to it.

Full instructions are in the official Idol repository.

The honest truth

Every tool has trade-offs. These are the ones the official documentation admits to, plus what we have not been able to check.

You get a splat, not a mesh

Idol writes Gaussian splats. The model card describes .ply splat output only and does not mention mesh export. If you need a textured mesh for a game engine, our Hunyuan3D 2.1 guide covers a tool built for that, with higher memory figures.

The "under 2 seconds" line has no hardware next to it

The card says Idol "can run in under 2 seconds" but does not say on which GPU or at which settings. The fastest row in its own timing table is 3.4 seconds on a B300 data center card at 50 steps. Plan around the table, not the headline.

The 7.2 GB peak comes without settings

The card does not say how many Gaussians or steps produced the 7.2 GB figure. Asking for 1,048,576 Gaussians may need more memory than the default 262,144. The card does not give a number for that, and we have not benchmarked Idol.

Two gated models stand between you and the first run

Idol stops before downloading anything if your Hugging Face account has not accepted the DINOv3 and FLUX.2-dev terms. Read those terms yourself: they are separate from Idol's license.

It is an NVIDIA story on desktop

The install line names "an NVIDIA GPU with CUDA", and the SDK README says the fast sampling path is CUDA-only. The same README reports that an M4 MacBook Air takes about 18 minutes for the default settings and needs 24 GB. AMD and Intel cards are not documented, so the chart, picks and price table cover NVIDIA GeForce RTX cards.

The license is permissive, with a footnote

The model card lists Idol under Apache-2.0, and the SDK repository carries the Apache License 2.0 text. The two gated encoders it downloads have their own terms, so check all three before commercial use.

Idol questions

How much VRAM does Idol need?

The official model card says Idol runs on consumer GPUs using 7.2 GB of VRAM. Its component table lists about 3.7 to 4.4 GB of weights and a 7.2 GB peak for a full request, so an 8GB card clears the published figure with little headroom.

Can Idol run on an 8GB graphics card?

By the model card figures, yes: the full-request peak is 7.2 GB. Running the DINOv3 encoder in fp32 uses 1.7 GB more memory and the card gives no total for that mode, so keep the default bf16 setting on an 8GB card.

Does Idol make a mesh I can use in a game engine?

No. Idol writes 3D Gaussian splats as a standard 3DGS .ply file, which opens in Gaussian splat viewers. The model card does not mention mesh export.

My picks

If this guide has you shopping, here is where we would start. Each card meets one of the official Idol tiers above. Prices are the lowest New and Used listings we have recorded on Amazon US for that model, not live checkout prices.

  • Budget pick

    RTX 4060 8GB+

    For trying image-to-3D on a small budget. Its 8GB clears the 7.2 GB peak in the model card, with little to spare, so keep the default settings.

    Lowest New: $599.99 New, 8GB, recorded 2026-11-05 06:55:40 UTC

    Lowest Used: $459.99 Used, 8GB, recorded 2026-11-03 23:32:11 UTC

    See every RTX 4060 listing

  • Sweet spot

    RTX 3080 10GB+

    The one mid-range card the authors timed themselves: 46.2 seconds per request in the model card's table. It is an older card, so compare Used listings.

    Lowest New: $899.00 New, 10GB, recorded 2026-11-19 07:29:16 UTC

    Lowest Used: $449.99 Used, 10GB, recorded 2026-09-29 16:22:32 UTC

    See every RTX 3080 listing

  • No-compromise

    RTX 4090 24GB+

    For shorter waits and room for other 3D tools. The model card times it at 19.1 seconds per request, and 24GB also meets the middle tier in our Hunyuan3D 2.1 guide.

    Lowest New: $4,990.00 New, 24GB, recorded 2026-11-21 07:08:36 UTC

    Lowest Used: $3,299.99 Used, 24GB, recorded 2026-09-29 16:19:42 UTC

    See every RTX 4090 listing

No pressure, though. The best card is often the one already in your PC, so check it against the chart before you buy anything.

Lowest recorded prices by tier

Want the cheapest way into each tier, whatever the model? These are the cheapest New and Used listings we have recorded among NVIDIA GeForce RTX cards with at least that much VRAM. Used includes renewed and refurbished cards. Each is a recorded Amazon US price, not a live checkout price; the model page shows every listing and its condition.

Computed from recorded Amazon US listings when this page loads.
TierLowest NewLowest Used
Minimum (7.2GB+)$419.99 RTX 5050 8GB
Recorded 2026-10-24 11:09:56 UTC
$249.97 RTX 2060 SUPER 8GB
Recorded 2026-09-30 05:03:30 UTC
Comfortable (10GB+)$459.99 RTX 3060 12GB
Recorded 2026-11-05 19:57:25 UTC
$419.97 RTX 3060 12GB
Recorded 2026-11-21 07:09:48 UTC
Best (24GB+)$2,499.99 RTX 3090 24GB
Recorded 2026-11-21 07:11:12 UTC
$1,777.22 RTX 3090 24GB
Recorded 2026-10-24 04:46:46 UTC

Compare cards

Estimate memory for language models with the LLM VRAM estimator, see typical prices per memory size in GPU prices by VRAM, or compare pairs: RTX 3060 vs RTX 4060, RTX 4090 vs RTX 5090. More local AI GPU guides.

Sources

We have not benchmarked these tools. Requirements change between releases, so check the repository before you buy.