Model zoo

Both load with AutoModel.from_pretrained. Meta's own facebook/sam3 is a conversion source, not a nobg checkpoint — see Starting from Meta's weights.

FeyNobg

A BiRefNet with a Swin backbone, trained for dichotomous image segmentation. It is the default: it takes an image and nothing else, and it predicts the matte at full input resolution, so soft edges — hair, fur, motion blur — come out as genuinely soft alpha.

from nobg import AutoModel

model = AutoModel.from_pretrained("feyninc/FeyNobg")
model.process("input.jpg").save("output.png")

See the BiRefNet reference for the config fields and methods.

MultiMatte

The SAM 3 architecture — a promptable, open-vocabulary detector — trained for matting. nobg wraps the Apache-2.0 transformers implementation and collapses its prompt-conditioned segmentation into a single alpha matte, so it drops into the same flow as FeyNobg, and gains the capability BiRefNet does not have: choosing what to cut out.

from nobg import AutoModel

model = AutoModel.from_pretrained("feyninc/multimatte")
model.process("input.jpg", "the dog").save("dog.png")

Prompt-free, it behaves like any other background remover — the config's default_prompt ("the main foreground subject") is supplied for you:

model.process("input.jpg").save("output.png")

The class is Sam3; AutoModel resolves it from the repo's nobg-sam3 tag. See Text and box prompts and the Sam3 reference.

Starting from Meta's weights

To convert Meta's untrained-for-matting checkpoint yourself instead, from_origin reads the upstream layout:

from nobg import Sam3

model = Sam3.from_origin("facebook/sam3")

Which one to reach for

The SAM3 architecture finds the right subject — measured on facebook/sam3, its matte agrees with FeyNobg at IoU 0.98 on a test photo — but it is a detector at heart: masks are predicted at a fraction of the input resolution and upsampled, so edges stay softer. On two test photos, 19–29 % of pixels land at intermediate alpha, versus 3 % for FeyNobg.

• MultiMatte when you need to choose what to cut out, by name or by box.
• FeyNobg when you need hair-level edges.

Loading a checkpoint

AutoModel reads the repo's Hub tags and returns the concrete class; AutoProcessor reads preprocessor_config.json (falling back to the model config) and returns the matching processor:

from nobg import AutoModel, AutoProcessor

model = AutoModel.from_pretrained("feyninc/FeyNobg")
processor = AutoProcessor.from_pretrained("feyninc/FeyNobg")

Concrete classes work too, if you would rather be explicit:

from nobg import BiRefNet, BiRefNetImageProcessor, Sam3, Sam3Processor

model = BiRefNet.from_pretrained("feyninc/FeyNobg")
processor = BiRefNetImageProcessor.from_pretrained("feyninc/FeyNobg")

model = Sam3.from_pretrained("feyninc/multimatte")
processor = Sam3Processor.from_pretrained("feyninc/multimatte")

Constructing from scratch (random init) uses the config dataclass:

from nobg import BiRefNet
from nobg.birefnet.modeling_birefnet import BiRefNetConfig

model = BiRefNet(BiRefNetConfig(image_size=512, embed_dim=128))