RMBG-3.0 (1038lab ComfyUI Quantized)

Quantized single-file safetensors (FP8 / NVFP4 / BF16) of Bria RMBG-3.0 for ComfyUI.

This repository hosts quantized single-file .safetensors weights of briaai/RMBG-3.0 formatted for native ComfyUI (Fibo2 / Load Diffusion Model).

RMBG-3.0 vs RMBG-2.0

Feature RMBG-2.0 RMBG-3.0
Architecture BiRefNet (single-pass segmentation) Fibo 2.0 Edit Turbo (4-step Diffusion Transformer)
Output Type Segmentation mask Continuous 8-bit alpha matte (true soft transparency)
Hard Cases Good on general subjects Much stronger on fine hair/fur, glass, mesh/fences, motion blur, and cluttered backgrounds
VRAM / Size ~1 GB (fast, runs on 4 GB+ GPUs) 4.3 GB (NVFP4) / 7.6 GB (FP8) / 15.1 GB (BF16) + Fibo2 CLIP & FLUX.2 VAE
License CC-BY-NC-4.0 Bria FIBO / CC-BY-NC-4.0 (Non-Commercial)

Available Files

  • RMBG-3.0-fp8.safetensors: FP8 (float8_e4m3fn, 7.57 GB) β€” recommended for RTX 30/40/50 series.
  • RMBG-3.0-nvfp4.safetensors: NVFP4 (5.43 GB) β€” optimized for NVIDIA Blackwell (RTX 50 series).
  • RMBG-3.0-bf16.safetensors: Full BF16 single-file (15.14 GB) β€” lossless reference.

ComfyUI Usage

Place the selected .safetensors file into:

ComfyUI/models/diffusion_models/

Paired with Fibo 2.0 components in ComfyUI:

  • Diffusion Model: Load Diffusion Model (RMBG-3.0-fp8.safetensors or RMBG-3.0-nvfp4.safetensors)
  • Text Encoder: Load CLIP (qwen3vl_4b_bf16.safetensors, Type: fibo2) in ComfyUI/models/text_encoders/
  • VAE: Load VAE (flux2-vae.safetensors) in ComfyUI/models/vae/

License & Attribution

Downloads last month
85
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for 1038lab/RMBG-3.0

Finetuned
briaai/RMBG-3.0
Finetuned
(2)
this model