Remove Background (BiRefNet)

How it fits

Typical backends

Caveats

• I/O contractThe output arm is a single downstream-resolved choice fixed at deploy time. The RGBA-cutout arm always emits four-channel Image and keeps the raw soft alpha; the mask arms instead binarise the matte at threshold, which has no effect on the cutout arm.
• AccuracyThe matte is predicted at inference_size and bilinearly resampled back to the source resolution, so edge sharpness scales with that size at proportional GPU cost; the model targets a single salient subject rather than every foreground object.
• I/O contractGreyscale input is broadcast to three channels before normalisation and any existing alpha channel is dropped on the way in.
• State lifetimeThe BiRefNet checkpoint loads offline from the bound model artifact (a transformers snapshot directory holding config.json, the safetensors weights and the repo's own modeling code) with `local_files_only`; `trust_remote_code` runs the snapshot's modeling code, so no Hub fetch occurs. Bind a BiRefNet checkpoint to the model slot at deploy.
• CostInference runs in float16 on GPU (BiRefNet's deformable convolution has no bf16 CUDA kernel) and float32 on CPU; CPU inference at a large inference_size is extremely slow.
• State lifetimeThe checkpoint, device, inference_size and threshold are captured once at startup; changing any of them requires a redeploy.