joeynyc/Studio-Image-Turbo-MLX-6bit-from-Qwen-2.1

Studio Image Turbo MLX — 6-bit

Built with Qwen. An unofficial MLX conversion of Qwen-Image-2.1-Turbo, using 6-bit transformer and text-encoder weights with MFLUX precision exclusions and a float32 VAE. No separate LoRA is used.

The original Qwen Research License applies: research/evaluation, non-commercial use only unless separately licensed by the rights holder. See LICENSE, NOTICE, and the retained upstream model card. Quantization does not grant additional usage rights.

Available exports: 4-bit · 6-bit (recommended).

Orbital Studio: native MLX output, 768 × 768, eight steps, seed 77. No retouching or upscaling. Prompt and generation settings.

Run on Apple Silicon

uv, Python 3.12, Git, and an Apple Silicon Mac are required for the commands below. This is an MFLUX diffusion model, not an MLX-LM language model or a drop-in Diffusers checkpoint.

uv venv --python 3.12
uv pip install --python .venv/bin/python huggingface-hub==1.33.0
.venv/bin/hf download joeynyc/Studio-Image-Turbo-MLX-6bit-from-Qwen-2.1 --local-dir ./studio-turbo-6bit
uv pip install --python .venv/bin/python -r ./studio-turbo-6bit/requirements-mac.txt
.venv/bin/python ./studio-turbo-6bit/generate.py \
  --prompt "A small red robot holding a white sign that says HELLO, clean studio photograph, soft light." \
  --seed 42 --width 512 --height 512 --output ./robot.png

Use the included runner: it applies the checkpoint's exact eight unshifted sigma steps, CFG 1, KV caching, staged component loading, and tiled VAE decode. The pinned upstream edit API does not automatically select this Turbo schedule. It saves the PNG and a .run.json timing/allocation report.

Conversion and validation

• Original checkpoint revision: d65dbc9a7e8f6b5479e33dee6030eaab2a906509.
• MFLUX commit: 76f8aeb3f1f1dbccffca524a451f79326bd794f7; MLX 0.32.2, MFLUX 0.20.0.
• Both original-precision MLX and official BF16 Diffusers were checked on two prompts with identical starting noise. Final latent cosine similarity was 0.99983 (robot) and 0.99995 (steak); images were visually near-identical.
• This 6-bit export ran natively on a 24 GB Apple M4 Mac at 512×512, eight steps: 54.13 s total, 6.34 GiB peak MLX allocation, zero swap growth. These are individual smoke measurements, not a general benchmark. MLX allocation is not total system memory.
• 6-bit is recommended for fidelity when memory/storage allow. Same-noise final latent cosine versus official BF16: 4-bit 0.97416/0.89965 (robot/steak), 6-bit 0.99952/0.99907. This is numerical agreement, not an aesthetic quality score.
• Image editing and arbitrary prompts are not qualified by these text-to-image smoke checks. Seeds across CUDA/Metal do not guarantee identical noise or pixels.
• Weight headers include modification/license notices. Packaging preserves every tensor payload byte from the tested export. conversion.json records payload hashes; SHA256SUMS records published file hashes.

Additional native resolution tests: 768×768 took 118.89 s with 7.37 GiB MLX peak and zero swap growth; 1024×1024 took 220.51 s with 8.45 GiB peak and 42.19 MiB system-wide swap growth. These larger tests used separate workloads and do not establish broad image quality.