nobg
API reference

AutoModel & AutoProcessor

Tag-based dispatch to the concrete model and processor classes.

from nobg import AutoModel, AutoProcessor

AutoModel.from_pretrained

AutoModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)

Reads the repo's Hub tags via model_info, then delegates to the concrete class's own from_pretrained with every argument untouched.

TagsClassPublished checkpoint
nobg-birefnet, birefnetBiRefNetfeyninc/FeyNobg
nobg-sam3, sam3Sam3feyninc/multimatte (MultiMatte)

Anything else raises ValueError("this model is not part of nobg").

Needs the network

Tag lookup is a Hub API call, so AutoModel cannot resolve a purely local directory. Use the concrete class (BiRefNet.from_pretrained("./my-model")) for offline loads.

AutoProcessor.from_pretrained

AutoProcessor.from_pretrained(pretrained_model_name_or_path, **kwargs)

Reads preprocessor_config.json (or the nested image_processor block of processor_config.json) and dispatches on the recorded image_processor_type, then re-loads through that class's own from_pretrained so all keyword arguments are handled the way the class expects.

image_processor_typeClass
BiRefNetImageProcessorBiRefNetImageProcessor
Sam3ProcessorSam3Processor
Sam3ImageProcessor, Sam3ImageProcessorFastSam3Processor

The last two are upstream type names: a raw transformers SAM3 repo such as facebook/sam3 records them, and nobg's Sam3Processor is a drop-in superset, so they map onto it rather than raising.

Fallback for repos without a processor config

If the file is missing (legacy BiRefNet checkpoints), AutoProcessor falls back to the Hub tags plus config.json:

  • BiRefNet → BiRefNetImageProcessor(size=config.image_size), defaulting to 1024
  • SAM3 → Sam3Processor with a size-only image processor (defaulting to 1008), the repo's own tokenizer, and default_prompt from the config

Model tags

Models self-declare their tags at class-definition time, which is what makes both dispatchers work after a push_to_hub:

class BiRefNet(
    nn.Module,
    Revised_Mixin,
    library_name="nobg",
    repo_url="https://github.com/feyninc/nobg",
    tags=["nobg", "nobg-birefnet"],
    ...
):

Every model carries "nobg" plus one unique model tag. See Contributing for the rules a new model has to follow.

ONNX export

Export the matte to ONNX, push it, and run it under onnxruntime with the same API.

BiRefNet

The bilateral reference matting model — config, forward, and its methods.

On this page

AutoModel.from_pretrainedAutoProcessor.from_pretrainedFallback for repos without a processor configModel tags