RMBG-1.4

RMBG-1.4 is BRIA AI's open-source image background-removal model v1.4, first released in March 2024, licensed under Apache-2.0 (commercial-use OK). The GitHub repository is a mirror maintained by chrisk3140; the model weights live on HuggingFace as briaai/RMBG-1.4 (100k+ downloads). Built on the BiRefNet architecture with 1.4GB parameters, it fine-tunes ISNet / U²-Net / SAM specifically for products and portraits, delivering significantly better edge detail (hair strands, semi-transparent objects, product edges) than previous-generation open-source options.

Core capabilities center on commercial-grade matte quality + independently distributable weights. Architecture: BiRefNet (Bilateral Reference Network), which uses bidirectional reference maps and a high-frequency detail recovery module to produce high-quality alpha masks; on the BF1.4 test set (portraits / products / transparent objects) it exceeds 95% IoU. Output: single-channel alpha mask (grayscale PNG), optional composite (original + mask + optional background). Inference: PyTorch + ONNX Runtime + CoreML; 1024x1024 takes ~2-4s on CPU, 5-10x faster on GPU (CUDA / CoreML / MPS); FP16 halves VRAM. Deployment: HuggingFace Transformers one-line load + infer; ComfyUI has a RMBG-1.4 node; Docker / onnxruntime-server for self-hosting. BRIA AI commercial backing: BRIA AI is a B2B-grade image generation / processing vendor, and RMBG is their open-source line for enterprise customers. Quality is comparable to commercial offerings like remove.bg and PhotoRoom, but the commercial license is unambiguous (Apache-2.0).

Use cases include: commercial product image (e-commerce / advertising / marketing) production-environment background removal (commercial-use license + high quality), AIGC training data mask annotation, media post-production transparent-bottom processing, and designer material compositing. Hardware floor: runs on CPU (ONNX Runtime, slow but usable); GPU recommended (8GB+ VRAM); FP16 drops to 4GB VRAM. Onboarding: pip install transformers torch pillow → load briaai/RMBG-1.4 → inference; ComfyUI users install the ComfyUI-BRIA-RMBG node.

Licensed under Apache-2.0, commercial use + patent grant + trademark disclaimer — suited for commercial product integration; the explicit patent grant is the compliance advantage versus U²-Net (Apache-2.0 same tier) and ISNet (MIT). Positioning versus VisLane's already-listed backgroundremover and rembg (pending): rembg is a "multi-model unified interface library" (covering U²-Net / ISNet / BiRefNet / SAM); RMBG-1.4 is "single model maximally optimized" (BRIA AI fine-tuned BiRefNet). rembg uses RMBG-1.4 as an optional backend (model name briaai-rmbg), but using RMBG-1.4 standalone gives finer tuned parameters and clearer commercial license docs.