Eraser

A small background-removal model (3.9 M parameters, 15.7 MB as fp32 ONNX, 4 MB as int8) distilled from BiRefNet. Give it an RGB photo, get back an alpha matte. It runs on CPU in roughly 50-100 ms per 512Γ—512 image, so it fits browsers, mobile and batch pipelines where a 200+ MB model is too heavy.

Intended use: cutting out the main subject of product photos (catalog / e-commerce) and everyday photos, including portraits, with a clear subject. It is not designed for complex scenes with thin structures or ambiguous subjects (e.g. the DIS5K benchmark), nor for picking one specific object out of a scene with several (see the Finegrain results below).

Status: a compact student. On portraits and natural photos it keeps about 93-98 % of BiRefNet's agreement with human masks at 1/56 of the size (see Benchmarks); it is clearly weaker on busy scenes, ambiguous subjects and thin structures. See Limitations.

Files

File Size What it is
onnx/model.onnx 15.7 MB fp32, reference
onnx/model_fp16.onnx 7.9 MB fp16 weights, fp32 inputs/outputs. Same output as fp32 (see below)
onnx/model_int8.onnx 4.1 MB int8 static quantization (QDQ). ~2Γ— faster on CPU, small quality drift
pytorch/model.safetensors 15.7 MB fp32 weights for PyTorch
model.py architecture definition (needed for the PyTorch weights only)
examples/inference.py ONNX Runtime example: image β†’ RGBA cut-out
benchmark/ benchmark script, instructions and raw results (see Benchmarks)
NOTICE.md data and teacher attribution

Input / output

  • Input image: float32 [batch, 3, 512, 512], RGB, values in [0, 1]. Do not apply mean/std; ImageNet normalization is inside the model. Resize the whole image to 512Γ—512 (plain bilinear, no letterboxing).
  • Output alpha: float32 [batch, 1, 512, 512] in [0, 1]. Resize it back to the original size and use it as the alpha channel.
  • Batch size is dynamic; the spatial size is fixed at 512Γ—512.

Quick start (ONNX Runtime)

import numpy as np, onnxruntime as ort
from PIL import Image

sess = ort.InferenceSession("onnx/model.onnx", providers=["CPUExecutionProvider"])

img = Image.open("photo.jpg").convert("RGB")
w, h = img.size
x = np.asarray(img.resize((512, 512), Image.BILINEAR), dtype=np.float32).transpose(2, 0, 1)[None] / 255
alpha = sess.run(None, {"image": x})[0][0, 0]                          # (512, 512) in [0, 1]
alpha = Image.fromarray((alpha * 255).astype("uint8")).resize((w, h), Image.BILINEAR)

img.putalpha(alpha)
img.save("cutout.png")

Or run python examples/inference.py photo.jpg --model onnx/model_int8.onnx.

PyTorch

import torch
from model import load_model

model = load_model("pytorch/model.safetensors")        # eval mode
with torch.no_grad():
    alpha = model(torch.rand(1, 3, 512, 512))           # (1, 1, 512, 512) in [0, 1]

Precision variants

Measured against the fp32 model on 100 photos that were not used for calibration or training (Open Images validation split, resized to 512Γ—512). CPU time is batch 1 on one desktop CPU (your numbers will differ).

Variant Size CPU time mean abs. diff vs fp32 IoU vs fp32 (Ξ±>0.5) 95th-percentile image (mean diff)
fp32 model.onnx 15.7 MB 93 ms (reference) (reference) (reference)
fp16 model_fp16.onnx 7.9 MB 95 ms 0.0001 0.9996 0.0004
int8 model_int8.onnx 4.1 MB 50 ms 0.0110 0.967 0.034
  • fp16 is effectively lossless. On a plain CPU it is not faster; its benefit is size, and speed on GPUs / NPUs with fp16 support.
  • int8 halves CPU time and shrinks the file 4Γ—, at the price of a small drift from fp32 (mostly on soft edges). The edge-refinement layers were kept in fp32. It was calibrated on 100 photos; if your images are very different (e.g. line art, X-rays), re-calibrate on your own data.

Benchmarks (real ground-truth masks)

Measured with the script in benchmark/ (included in this repo), on public datasets with human-made masks. The teacher (BiRefNet, 1024 px) was run on the same images with the same code, so the last column is a measured retention (student Γ· teacher). Student = fp32 model. S = S-measure, IoU at threshold 0.5 (higher is better), MAE lower is better; all at a long side of at most 1024 px.

Dataset What it is Images S teacher β†’ student IoU teacher β†’ student MAE teacher β†’ student Retention S / IoU
P3M-500 (NP) portraits, soft alpha 500 0.982 β†’ 0.959 0.993 β†’ 0.964 0.003 β†’ 0.017 97.7 % / 97.1 %
AIM-500 natural photos, soft alpha 500 0.967 β†’ 0.921 0.966 β†’ 0.899 0.006 β†’ 0.035 95.3 % / 93.1 %
DUT-OMRON everyday photos, main object 1,000 of 5,168 (random, fixed) 0.869 β†’ 0.789 0.774 β†’ 0.637 0.045 β†’ 0.103 90.8 % / 82.3 %
DUTS-TE everyday photos, main object 1,000 of 5,019 (random, fixed) 0.929 β†’ 0.812 0.870 β†’ 0.666 0.021 β†’ 0.085 87.4 % / 76.6 %
Finegrain Products e-commerce, only the product 120 0.786 β†’ 0.684 0.639 β†’ 0.482 0.090 β†’ 0.161 87.1 % / 75.4 %
DIS5K (DIS-VD) thin structures, hard scenes 470 0.912 β†’ 0.718 0.842 β†’ 0.518 0.031 β†’ 0.144 78.7 % / 61.6 %

The student is about 56Γ— smaller than the teacher (3.9 M vs ~220 M parameters; 15.7 MB vs 880 MB in fp32). Full tables (maxF, wF, E, all three precisions) are in benchmark/results/<dataset>/report.md.

How to read it:

  • Close to the teacher on portraits and natural photos (P3M-500, AIM-500: 93-98 % retention). This is the target use.
  • Noticeably behind on everyday saliency sets (DUTS-TE, DUT-OMRON: 77-82 % IoU retention), where the "main object" is often ambiguous.
  • Finegrain Products asks for only the product (the bottle in a hand, the dress without the model). Even the teacher scores only 0.64 IoU there, because it also cuts out the hand or the person. The student keeps 75 % of that. Plain packshots looked fine in spot checks (not measured separately).
  • DIS5K is a stress test (thin, complex structures; the teacher was trained on that distribution, the student was not). 62 % IoU retention.
  • Sanity check of the benchmark: the teacher gets S = 0.912 on DIS-VD here, matching the 0.911 published for BiRefNet.
  • Precision variants: fp16 is identical to fp32 and int8 is within 0.005 on every metric on every dataset.
  • The test sets were not used for training. Check each dataset's original license before reusing it.
  • Method notes: the student numbers come from a CPU run, the teacher numbers from a GPU run on Colab (L4, fp16 autocast). The two runs use the same code and the same fixed random subsets (image counts match). Teacher figures were taken from the notebook output at 3-decimal precision.

Agreement with the teacher (validation during training)

Measured against the teacher's masks (BiRefNet at 1024 px), not human ground truth, on images never used for training (an 80/20 split by product for ABO; held-out photos for Open Images). Lower MAE and higher IoU are better.

Validation set Images MAE IoU (Ξ±>0.5)
Product photos (ABO catalog) 11,435 0.036 0.926
Products composited on unseen backgrounds 11,435 0.032 0.923
General photos (Open Images: people, animals, objects in scenes) 5,157 0.102 0.746

The gap between product and general photos is real: general scenes are harder and the teacher's own "main subject" choice is ambiguous in many of them. A larger variant of this student (with an extra ResNet layer4, 13.7 M parameters) improved general-photo MAE to 0.088 / IoU 0.771 but is 3.5Γ— larger and ~1.9Γ— slower, so it is not part of this release.

Architecture and training

  • Student: ResNet18 encoder (up to layer3, ImageNet-initialized) β†’ pyramid pooling (bins 1/2/4/8) β†’ U-Net style decoder with deep supervision β†’ full-resolution edge refiner. 3.92 M parameters, input 512Γ—512.
  • Teacher: BiRefNet, 1024Γ—1024, fp16. Its soft alpha output is the training target ("pseudo-labels"); unreliable teacher masks (very uncertain, or nearly empty / nearly full) were filtered out.
  • Data: 60k product photos (Amazon Berkeley Objects), plus 30k general photos (Open Images) in a second stage; products were randomly pasted over real backgrounds (COCO, DTD) to avoid learning "white background = background".
  • Loss: BCE + IoU + 2Β·Sobel gradient loss (sharp edges), with auxiliary heads.
  • Stage 1: 30 epochs on products. Stage 2: 15 epochs fine-tuning on a 50 / 50 mix of products and general photos.

Limitations

  • Hard or ambiguous scenes. Cluttered backgrounds, large close-up portraits, dark painted backgrounds and scenes where "the main object" is unclear can produce soft grey halos or blobby silhouettes. Typical failure: faint haze over large background areas.
  • It imitates the teacher, including its mistakes and its notion of what the "main subject" is. Do not expect it to beat BiRefNet.
  • Fine structures (hair, fur, thin wires, transparent objects) are less precise than in the teacher. The output is 512Γ—512, then upscaled.
  • Fixed 512Γ—512 input. Aspect ratio is not preserved when resizing; very wide or tall images lose detail.
  • Human-labeled evaluation is limited to the public benchmarks above; the training-time numbers are agreement with the teacher.
  • Licensing is non-commercial (see below).

License and data

Released under CC BY-NC 4.0 (non-commercial), because training used Amazon Berkeley Objects, which is CC BY-NC 4.0. Other sources: Open Images (images CC BY 2.0), COCO and DTD (backgrounds), BiRefNet (teacher), torchvision ResNet18 (initialization). Details and links are in NOTICE.md. Check every source's terms before any commercial use.

Acknowledgements

  • BiRefNet (Zheng et al.), the teacher.
  • Amazon Berkeley Objects, Open Images, COCO and Describable Textures Dataset.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support