How to use from the
Use from the
Diffusers library
pip install -U diffusers transformers accelerate
import torch
from diffusers import DiffusionPipeline

# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Ouinche/RMBG-3.0-MLX-4bit", dtype=torch.bfloat16, device_map="cuda")

prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]

RMBG-3.0-MLX-4bit

DiffusionBear RMBG-3.0-MLX-4bit

Full standalone 4-bit MLX pipeline for Bria AI RMBG-3.0 (briaai/RMBG-3.0), quantized for Apple Silicon (16 GB+ unified memory) and used by DiffusionBear (coming in v0.5.1).

Contents

  • transformer/: BriaFibo2Transformer2DModel from briaai/RMBG-3.0, quantized to 4-bit (group_size=64, ~4.26 GB).
  • text_encoder/: 4-bit Qwen3-VL-4B-Instruct text/vision encoder.
  • vae/: FLUX.2 VAE encoder/decoder.
  • processor/, scheduler/, fonts/, model_index.json: full BriaFibo2Pipeline configuration.

Matte Resolution Presets & Measured M1 16 GB Timings

DiffusionBear exposes three matte-inference presets (4 denoising steps) plus an optional Target subject prompt (leave blank for automatic foreground removal, or specify e.g. "the woman on the left", "the dog"). The output PNG always retains the original source resolution; only the alpha matte inference resolution scales.

Preset Matte Inference Puppy 1024Γ—1024 (M1 16 GB) Edge Quality & Recommended Use
Fast (512 px) 512 px max side 163.81 s (~2m 44s) Quick previews, game sprites, and 512–768 px workflows; slightly softer fur/hair when upscaled to 1024 px.
HQ (768 px) 768 px max side 303.18 s (~5m 03s) Best speed/detail balance for 768–1024 px assets; preserves open-knit holes (342.06 s) and botanical stems (307.38 s).
Native (1:1) Full source size 603.49 s (~10m 03s) Maximum per-strand fur and whisker precision at native pixel resolution (~3.7Γ— slower than Fast 512 px).

Visual Samples (Official RMBG Test Images on M1 16 GB)

1. Terrier Puppy (1024Γ—1024: Input vs 512 px vs 768 px vs Native 1024 px)

Input (1024Γ—1024) Fast 512 px (163.81 s) HQ 768 px (303.18 s) Native 1024 px (603.49 s)
Input 512px 768px 1024px

2. Open-Knit Crochet Cardigan (768Γ—768, 342.06 s) & Wildflower Bouquet (768Γ—768, 307.38 s)

Crochet Input Crochet Cutout (RGBA) Bouquet Input Bouquet Cutout (RGBA)
Crochet Input Crochet RMBG Bouquet Input Bouquet RMBG

βš–οΈ License and Attribution

This model is a derivative of RMBG-3.0 / FIBO by Bria AI. The model weights are Β© BRIA AI and licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

Downloads last month
34
MLX
Hardware compatibility
Log In to add your hardware

4-bit

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

Model tree for Ouinche/RMBG-3.0-MLX-4bit

Finetuned
briaai/RMBG-3.0
Finetuned
(5)
this model