U²-Net, ISNet, and SAM: A Deep Technical Comparison of Three Cutout Models
Published: 2026-10-10
Why Model Choice Matters
Different segmentation models make different trade-offs between accuracy, speed, and resource consumption. A high-precision model that runs in 5 seconds on a server may be impossible to run in a phone browser. Choosing the right model directly determines the cutout experience users get.
Technical Overview of the Three Models
U²-Net
ISNet
SAM (MobileSAM)
Year released
2020
2022
2023
Parameters
4.7M (lite)
~44M
~10M
Training data
DUTS (10K)
HRSOD/LDIS
SA-1B (1.1B masks)
Inference speed*
~200ms
~800ms
~50ms (decoder)
Edge accuracy
Medium
Excellent
Depends on prompt
Runs in browser
Yes — 2.4MB
No — needs a server
Yes — 23MB quantized
*Approximate figures on a desktop-class CPU (Intel i7) with 320×320 input.
U²-Net: Lightweight All-Purpose Segmentation
U²-Net's core innovation is its nested U-shaped structure. In a traditional U-Net, each level of the network is a convolutional block; in U²-Net, each level is itself a small U-Net (called an RSU block), which enlarges the receptive field without inflating the parameter count.
Strengths
Tiny footprint (only 2.4MB after FP16 quantization), runs in real time in the browser via WebAssembly
Reliable segmentation of everyday subjects — people, products, animals
Fully automatic: no user prompts needed; it segments the most salient object on its own
Weaknesses
Less precise on complex edges (hair, transparent objects) than larger models
Fixed 320×320 input resolution, so edges can look jagged after upsampling
In multi-object scenes it can only pick the single most salient one
ISNet: High-Precision Edge Segmentation
ISNet focuses on edge quality. It introduces a "semantic edge attention" mechanism that explicitly supervises edge information at every level of the decoder, producing output masks that are noticeably sharper at object boundaries.
Strengths
Industry-leading edge accuracy, excellent with hair-level detail
Ranked first on DIS5K, a benchmark of hard segmentation cases
Supports higher-resolution input (1024×1024)
Weaknesses
Large model (~44M parameters), not practical to run in a browser
Requires a GPU server
SAM: Promptable Universal Segmentation
SAM (Segment Anything Model), released by Meta AI in 2023, was the first general-purpose segmentation model trained on a mega-scale dataset (SA-1B, containing 1.1 billion masks). Its breakthrough is being "promptable": click anywhere on the image and the model segments the corresponding object.
Strengths
Extremely versatile — one model segments everything (people, animals, furniture, food, and more)
Supports point, box, and text prompts
Excellent interactive refinement — click to erase or restore regions
Weaknesses
In fully automatic mode (no prompt), accuracy trails dedicated salient-object segmentation models
The full model is huge (~600MB); the quantized MobileSAM is about 23MB
Real-World Test: The Same Image Across Three Models
Simple objects (shoes, cups): all three perform similarly — U²-Net is enough
Human hair: ISNet clearly beats U²-Net; SAM gets close once correctly prompted
Transparent objects (glass): ISNet is the most stable; SAM needs several prompt corrections
Multi-object scenes: SAM can select the exact target; the other models only take the most salient object
Our Multi-Engine Architecture
JuanJuan Tools uses a multi-engine routing strategy that guarantees both availability and quality: