Idol API on fal: Image-to-3D Gaussian Splats for $0.10 per Generation

fal announced hosted access to Sperid Labs’ Idol on October 30, with image-to-3D generation priced at $0.10 per request. The model turns a single object image into a 3D Gaussian splat, giving developers a way to try the workflow without deploying the inference stack themselves.

The fal announcement concerns hosted availability, not the first release of the underlying model. The practical question is whether its output fits the application you are building.

Idol API pricing and output

The Idol endpoint on fal lists $0.10 per generation. At that listed rate, 100 generations would cost $10 before any separate application infrastructure costs.

The endpoint returns a .ply Gaussian splat and a GLB point-cloud preview. Its example response also includes the preprocessed image, camera information, Gaussian count and timing fields.

A GLB download does not automatically mean a conventional polygon mesh. Here, fal explicitly describes it as a point-cloud preview. That matters if your next step assumes editable surfaces, a clean topology or a mesh suitable for rigging.

For an interactive object preview, a splat may be the representation you want. For a production asset pipeline, the existence of a downloadable file is only the beginning of the compatibility check.

What Idol generates from one image

Sperid Labs’ inference SDK accepts photographs, renders and generated images. It removes the background, frames the object and produces a Gaussian representation, including inferred views that the source picture never showed.

The SDK documents a range of 32,768 to 1,048,576 Gaussians, with 262,144 as the default. Those are documented local-model settings, not a promise that every setting is exposed identically through fal.

The unseen surfaces are especially important for product use. A plausible back view is not evidence of the real object’s shape. A catalog application needs to decide whether it is offering a visual approximation or an accurate representation before presenting the result to customers.

Hosted access versus running Idol locally

Local deployment is available, but the SDK requires access to its DINOv3 and FLUX.2 dependencies. Sperid Labs reports 7.2GB peak VRAM usage and provides hardware-specific timings. Its own 50-step table lists approximately 6.7 seconds on an H100 and 19.1 seconds on an RTX 4090.

These are the developer’s measurements, not independent tests or a fal latency guarantee. Initial local inference can also be slower while compilation occurs.

Hosted pricing makes the first experiment easier to budget. The useful evaluation is still downstream: whether the generated object renders correctly in your viewer, preserves the details your users care about and saves enough work after review. A cheap generation is less valuable if the receiving workflow needs an entirely different kind of asset.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain product launches clearly, including the details that determine how developers actually use them.

Book a 30-minute call.