DiffusionBear
Coming in v0.5.1 FIBO 2.0 4-bit MLX Engine Non-Commercial CC BY-NC 4.0

Bria RMBG-3.0 on Apple Silicon

4-bit MLX port of Bria AI's RMBG-3.0 background-removal model (Ouinche/RMBG-3.0-MLX-4bit), running locally on the same sequential FIBO 2.0 engine inside DiffusionBear (coming in v0.5.1).

7.5 GB on disk · 4 steps · Outputs native RGBA PNG · Tested on MacBook Pro M1 16 GB

Official RMBG Sample Set · 4-bit MLX

Fur, open-knit fabric & foliage cutouts

Each pair below runs an official Bria RMBG sample image through Ouinche/RMBG-3.0-MLX-4bit on an Apple M1 (16 GB). Click any image to inspect full resolution.

Sample 1 · Terrier puppy — Matte resolution comparison (1024×1024)

The output PNG always keeps the full source resolution (1024×1024), while the alpha matte is inferred at 512 px, 768 px, or full 1:1 resolution.

Input · 1024×1024 RGB
Fast (512 px) · 163.81 s · ~2m 44s
HQ (768 px) · 303.18 s · ~5m 03s
Native (1:1 · 1024 px) · 603.49 s · ~10m 03s

Sample 2 · Open-knit crochet cardigan (768×768)

Input · 768×768 RGB
RMBG-3.0 (4-bit MLX) · 768×768 RGBA · 342.06 s

Sample 3 · Wildflower & fern bouquet (768×768)

Input · 768×768 RGB
RMBG-3.0 (4-bit MLX) · 768×768 RGBA · 307.38 s

Options & Speed / Quality Trade-off

Matte resolution presets & target subject

RMBG-3.0 lets you trade inference time against alpha-matte sharpness, and optionally isolate a specific subject in multi-subject scenes (e.g. "the woman on the left", "the dog") by typing a target description or leaving it blank for automatic foreground removal.

Preset Matte inference Puppy 1024×1024 (M1) Edge quality & recommended use
Fast (512 px) 512 px max side 163.81 s Quick previews, game sprites, and 512–768 px inputs; slightly softer fine hair/fur when upscaled to 1024 px.
HQ (768 px) 768 px max side 303.18 s 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 Maximum per-strand fur and whisker precision at native pixel resolution; ~3.7× slower than Fast (512 px).

Specs & Licence

Shared FIBO 2.0 4-bit MLX architecture

Runtime details

HF Repository
Ouinche/RMBG-3.0-MLX-4bit
Engine
FIBO 2.0 sequential MLX
Disk Footprint
7.5 GB (4-bit)
Default Steps
4
Output
Native RGBA PNG
Availability
DiffusionBear v0.5.1

Non-Commercial License (CC BY-NC 4.0)

Like Bria FIBO 2.0 Edit Turbo, RMBG-3.0 weights and the 4-bit MLX port are released under CC BY-NC 4.0 for non-commercial evaluation, personal use, and research. Commercial deployment requires a separate agreement with Bria AI.

Related: FIBO 2.0 Edit Lab · All model licences