8-bit vs 16-bit vs 32-bit models: which quality setting to choose

The Settings panel offers three versions of the background removal model: Fastest (8-bit quantized, default), Balanced (16-bit), and Best quality (full precision). They are the same model, trained once, stored at different numerical precision. This guide explains what that means, what we measured, and which one to choose.

What precision means

A neural network is mostly a very large collection of numbers, called weights, that are multiplied and added together to turn your image into a mask. How precisely those numbers are stored determines the file size and how much work it takes to compute with them.

• 32-bit (full precision, fp32): each weight is a standard 32-bit floating-point number. This is how the model was trained, so it is the reference version.
• 16-bit (half precision, fp16): each weight uses 16 bits. Half the size, with a small loss of precision that rarely matters for image tasks.
• 8-bit (quantized, q8): weights are converted to 8-bit integers plus scaling factors. About a quarter of the full size. Quantization is a standard technique for running models efficiently on CPUs and phones.

The three versions side by side

The download is a one-time cost: your browser caches the model, so later visits load it from disk. Switching versions downloads the new one once. The times above were measured with ONNX Runtime's native CPU engine (Apple M1 Pro, 10 cores, 32GB RAM); the browser version is slower in absolute terms, but the ratio between the versions is similar.

How we measured the difference

We ran all three versions on the six example photos used across this site, then compared each mask with the full-precision mask pixel by pixel. Alpha runs from 0 (transparent) to 255 (opaque); a difference above 32 (about 12%) is roughly where it becomes visible on a contrasting background.

The 16-bit model matched full precision almost exactly on every photo. The 8-bit model was within one alpha level on average for most photos; the visible differences were concentrated in two places: the fine strands of the curly hair portrait, and the semi-transparent body of the glass bottle. The full tables, including speed and file sizes, are in our test results.

Note that on a CPU the 16-bit version was not faster than the 32-bit one: CPUs compute 16-bit weights by converting them to 32-bit. Its advantages are the smaller download and better speed with WebGPU on a graphics card.

This is a measurement on six photos, not a guarantee for every image. Very fine detail, such as flyaway hair against a busy background, and glass are where you are most likely to notice a difference. If a result is close but not quite right, trying the 16-bit version is a quick experiment.

Which one to choose

Use 8-bit (default) when

• You are on a phone, tablet, or a computer with limited memory. On these devices, the tool only offers the 8-bit model and CPU processing, to stay within the browser's memory limits.
• You are on a slow or metered connection.
• You process product photos with clear outlines, where all three versions produce the same result.
• You are running large batches and want the fastest throughput.

Try 16-bit when

• A portrait with loose hair or a long-haired pet is not quite right with the default.
• You have a reasonably fast connection and a desktop or laptop.
• You use WebGPU: GPUs handle 16-bit numbers efficiently, so the speed penalty is smaller. See WebGPU vs WebAssembly.

Use 32-bit when

• You want the reference result for a single important image, such as a cover photo, and do not mind waiting.
• You are comparing results and want to rule out precision as a cause.

For most people, 32-bit is not worth the larger download and slower processing: the difference from 16-bit is usually invisible.

Precision is not the main lever for quality

If a cutout has a problem, changing the model version is rarely the biggest improvement. In order of impact:

1. A better photo: sharper, better lit, with more contrast against the background. See shooting for clean cutouts.
2. Cropping closer so the subject fills more of the 1024 × 1024 image the model sees.
3. Select object to tell the model exactly what to keep.
4. The edge slider to tighten halos or recover clipped details.
5. Model precision, last.

Fixing halos and rough edges walks through these by symptom.

Summary

The three settings are the same model at different numerical precision. The 8-bit default is about a third of the 16-bit download, was about 35% faster on a CPU in our tests, and differed visibly on under 1% of pixels; 16-bit is worth trying for difficult hair and fur on a computer; and 32-bit is rarely needed. A better photo, a closer crop, and the selection and edge tools matter more than precision.