SceneWorks/boogu-image-mlx

Boogu-Image-0.1 — native MLX

Pre-converted, native MLX weights of Boogu-Image-0.1 for Apple Silicon, loaded by SceneWorks' mlx-gen Rust engine — no PyTorch / Python at inference. Boogu is a Lumina-Image-2.0 / OmniGen2-lineage flow-matching image model: a mixed single/double-stream DiT (~10.3B) + a Qwen3-VL-8B condition encoder + the FLUX.1 16-channel VAE.

Variants

Each subfolder is a complete snapshot — transformer/ (DiT) + mllm/ (Qwen3-VL encoder + tokenizer)

• vae/ — loadable directly.

Folder	Source	Task
base/	Boogu-Image-0.1-Base	text-to-image (true-CFG)
turbo/	Boogu-Image-0.1-Turbo	text-to-image (DMD few-step, no CFG)
edit/	Boogu-Image-0.1-Edit	reference image + instruction → edited image
base-bf16/ turbo-bf16/ edit-bf16/	(above)	same models, full precision

The bare-named folders are the default Q8 build; the -bf16 folders are the unquantized originals for maximum quality / experimentation.

Quantization (Q8 folders)

Group-wise affine Q8, group size 32 — the DiT hidden size (3360) is divisible by 32, not the usual 64. Quantized: the DiT Linears + the Qwen3-VL text-tower Linears. Kept full precision: the FLUX.1 VAE (decode-precision-sensitive), the Qwen3-VL vision tower (runs f32), and the token-embedding table.

Footprint (measured @ 1024²): Q8 ~35.5 GB peak → recommended 64 GB Mac. (Q4 would be ~27–30 GB / 48 GB; not shipped here.) bf16 runs require more.

Usage

Loaded by mlx-gen-boogu via BooguPipeline::from_snapshot("<variant folder>") (text-to-image, Turbo few-step, and single-reference image edit). Used by SceneWorks Image Studio.

License

Apache-2.0, inherited from the upstream Boogu/Boogu-Image-0.1-{Base,Turbo,Edit} checkpoints. These are re-quantized/repackaged copies of those weights for MLX; all credit to the Boogu authors.