mlx-community/Boogu-Image-0.1-Base-8bit

MLX (int8) conversion of Boogu-Image-0.1-Base (Apache-2.0) for Apple Silicon: bilingual (EN/ZH) text-to-image. OmniGen2-lineage pipeline (DiT + FLUX.1 VAE + FlowMatchEuler scheduler). The Qwen3-VL-8B-Instruct text encoder is the stock model (bit-identical) — referenced from Qwen/Qwen3-VL-8B-Instruct, not re-hosted.

int8 (group_size=32, attn+FFN Linears; embeds, time, norm_out and AdaLN kept bf16), the same layout as Turbo-8bit. Quantized from the bf16 tier; per-pass DiT cosine vs bf16 0.999897. ~12.5 GB DiT. Quant auto-detected via transformer/quant_config.json.

Memory (M5 Max, process phys_footprint, Swift package): 12.9 GB resident; peak 29.5 GB at 1024² (the Qwen3-VL conditioner is loaded per request and evicted before the denoise).

Use — Swift (MLXEngine)

boogu-image-swift ≥ 0.5.0: BooguImageConfiguration.base(quant: .int8) materializes this repo and the conditioner into the engine's store.

Use — Python

pip install mlx mlx-vlm
git clone https://github.com/xocialize/boogu-image-mlx && cd boogu-image-mlx && pip install -e .

from boogu_image_mlx.pipeline_mlx import BooguImagePipeline
from PIL import Image
pipe = BooguImagePipeline.from_pretrained("<this repo dir>", "Qwen/Qwen3-VL-8B-Instruct")
img = pipe.generate("a red panda surfing on a wave, photorealistic", height=1024, width=1024, steps=30, guidance=3.5)

Code: https://github.com/xocialize/boogu-image-mlx · https://github.com/xocialize/boogu-image-swift