LocalMuseAI/mlx-fibo-2-edit-turbo-4bit

Fibo 2 Edit Turbo — LocalMuse native INT4 package

An independent Swift/MLX conversion for LocalMuse's FiboRuntime. This is a downloadable runtime package, not a Diffusers or mlx-lm checkpoint. It supports one reference image and an edit instruction, 1024 × 1024, four steps, CFG 1. It does not offer text-to-image, masks, strength or multi-reference editing.

Original model: Bria AI Fibo 2.0 Edit Turbo, revision 7b133d62d92e897f1468af0427a73cc41a03cb5d. https://huggingface.co/briaai/Fibo-2.0-edit-turbo

Model weights © Bria AI, CC BY-NC 4.0. Non-commercial use only under that license; commercial use requires a separate agreement with Bria. See LICENSE for the complete terms and https://platform.bria.ai/licensing/pricing?model=fibo-2. This conversion does not grant commercial rights or imply Bria's endorsement. Qwen3-VL originates from Alibaba Cloud / Qwen Team; architecture code follows Apache-2.0 Transformers and Diffusers. See NOTICE for runtime attribution.

Changes: DiT attention/FFN and the first 35 Qwen text layers use affine INT4, group size 64. Sensitive DiT and vision tensors remain BF16. Text embedding is split into 4096-row files. Fibo's own original VAE is converted to native FP32 layout. Blocks load, execute and unload sequentially. Gist conditioning is cached; reference hidden states still run jointly at every denoising step. The source, conversion manifests and per-file SHA-256 values are included.

The original Diffusers pipeline was validated before conversion. Independent dequantization checks, synthetic layer/scheduler tests and a complete native Mac edit were performed with Metal API validation. Native Mac 1024-square full Edit took 384.96 seconds and peaked at 3.80 GB process memory (Debug, M1 Pro, 16 GB). Physical iPhone 17 Pro Max (iOS 27.2, Debug, Metal validation) completed the full Edit in 202.95 seconds, peak process footprint 3,861,727,576 bytes, peak MLX 3,737,646,208 bytes, estimated process limit 6,442,450,944 bytes. LocalMuse initially requires the nominal 12 GB device memory class. This package is about 8.21 GB on disk; model bytes are not all resident simultaneously.

The runtime uses paired three-axis RoPE, per-token clean/noisy conditioning, Qwen3-VL DeepStack and six normalized feature bundles, and the exact original four-step FlowMatch Euler schedule. Full-frame VAE attention/GroupNorm are preserved while convolution work is bounded in row bands.

LocalMuse project: https://github.com/cybsom/localmuse. Conversion provenance and authenticated file digests are included in this package. An editable mask is not supported by this port, and preservation of unrequested details is model behavior rather than exact pixel locking.