Ouinche/FIBO2-Edit-Turbo-MLX-4bit

FIBO 2.0 Edit Turbo - 4-bit MLX Native for Apple Silicon

FIBO 2.0 Edit Turbo (4-bit MLX) is a high-efficiency Apple Silicon native port of Bria AI's FIBO 2.0 Edit Turbo using Apple's MLX framework. Native GUI support is available in DiffusionBear v0.4.2.

By quantizing the 44-layer DiT Transformer and the Qwen3-VL-4B-Instruct text encoder into 4-bit affine format (group_size=64, bits=4), the total package on disk is reduced from 24.3 GB down to ~7.3 GB, and runtime peak memory drops from 17.4 GB (which crashes or causes severe disk swap on 16 GB Macs) down to < 6.8 GB.

It runs 100% in RAM with ZERO SSD swap on Apple Silicon Macs with 16 GB unified memory.

---

⚡ Empirical Benchmarks (Apple Silicon)

Memory Footprint & Compression

Component	Original Precision (bf16)	4-bit MLX (group_size=64)	Reduction
DiT Transformer (44 layers, dim 3584)	15.32 GB	4.26 GB	-72.2%
Text Encoder (Qwen3-VL-4B-Instruct)	8.27 GB	2.88 GB	-65.2%
VAE (FLUX.2 AutoencoderKL fp16)	0.34 GB	0.19 GB	-44.1%
Total Disk Storage	24.3 GB	7.33 GB	-69.8%
Runtime Peak VRAM	17.4 GB (OOM / Thrash)	6.8 GB	Fits in 16 GB Unified RAM

Generation Latency (4 Turbo Steps)

Resolution	Device	Steps	Cold Load (1st run)	Steady State (warmed)
512 × 512	Apple M1 (16 GB)	4	~73.9s	~58.8s (14.7s / step)
1024 × 1024	Apple M1 (16 GB)	4	~331s	~303s (75.9s / step)
1024 × 1024	Apple M4 / M5 (projected)	4	~35s	~25-30s (~7s / step)

---

🚀 Quickstart & Standalone Usage

1. Requirements

Install standard Apple Silicon MLX requirements:

pip install mlx mlx-vlm transformers huggingface_hub pillow

2. Run Inference in One Command

A. Text-to-Image Generation (4 Turbo steps)

# Clone the repository
git clone https://huggingface.co/Ouinche/FIBO2-Edit-Turbo-MLX-4bit
cd FIBO2-Edit-Turbo-MLX-4bit

# Fast preview (512x512)
python fibo2_generate.py "a red vintage bicycle leaning against a whitewashed wall on a narrow street" --width 512 --height 512

# High quality 1024x1024
python fibo2_generate.py "a cinematic shot of a majestic snow leopard perched on a rocky mountain peak at sunset, golden hour light, highly detailed" --width 1024 --height 1024 --steps 4 --seed 4242

B. Multimodal Image Editing (Reference Image)

Modify an existing image by passing --image:

python fibo2_generate.py "replace the jacket with a red leather one" \
  --image portrait.png --width 768 --height 512 --steps 4

C. Masked Inpainting Edit (Precise Local Regeneration)

Regenerate only the painted area by passing a binary mask (white = region to regenerate, black = untouched):

python fibo2_generate.py "replace the suitcase with a vintage brown leather one" \
  --image suitcase.png \
  --mask suitcase_mask.png \
  --composite \
  --width 768 --height 512 --steps 4

Note: Under the hood, masked inpainting conforms to the official FIBO 2.0 specification: the masked area of the reference image is mapped to flat mid-grey (RGB 128, 128, 128) before Qwen3-VL and VAE latent encoding, guaranteeing clean scene understanding without artificial boundary artifacts.

D. Multi-Reference Composition (Up to 4 Images)

python fibo2_generate.py "a portrait of the character in <image_1> in the art style of <image_2>" \
  --images person.png painting.png --width 512 --height 768 --steps 4

---

🧠 Staged Execution Architecture

To guarantee zero memory thrashing on 16 GB Apple Silicon devices, fibo2_generate.py executes in three sequential stages with explicit cache clearing:

1. Stage 1 (Text Encoder): Loads Qwen3-VL-4B-Instruct in 4-bit, extracts multi-layer hidden states (1, 6, L, 7680), and immediately evicts the text encoder from RAM.
2. Stage 2 (Transformer Denoising): Loads the 4.26 GB 4-bit DiT Transformer, runs the 4-step Turbo schedule, and evicts the transformer from RAM.
3. Stage 3 (FLUX.2 VAE Decode): Loads the lightweight 189 MB VAE decoder, decodes the latents into 24-bit RGB PIL image, and outputs PNG + JSON metadata.

---

📄 License & Attribution

• Weights License: © BRIA AI, licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). For commercial licensing, refer to BRIA AI licensing.
• MLX Pipeline Reference: Based on the MLX pipeline implementation by lvladikov.
• Quantization & Packaging: Packaged and maintained by Ouinche for the Apple Silicon MLX community and DiffusionBear.

Citation

@article{gutflaish2025generating,
  title={Generating an Image From 1,000 Words: Enhancing Text-To-Image With Structured Captions},
  author={Gutflaish, Eyal and Kachlon, Eliran and Zisman, Hezi and Hacham, Tal and Sarid, Nimrod and Visheratin, Alexander and Huberman, Saar and Davidi, Gal and Bukchin, Guy and Goldberg, Kfir and others},
  journal={arXiv preprint arXiv:2511.06876},
  year={2025}
}