Local text-to-image pipeline for Apple Silicon: Qwen-Image-2.1 on MLX with a Qwen3 safety guard, prompt rewriting and best-of-N reranking. CLI + Streamlit UI.
This repository contains no model weights. Models are downloaded separately and keep their own licenses. The table is generated from models.toml (tests/test_licenses.py keeps it in sync):
| Role | Model | Revision | License |
|---|---|---|---|
| guard | mlx-community/Qwen3Guard-Gen-0.6B-MLX |
919a4577 |
Apache License 2.0 |
| rewrite | mlx-community/Qwen3-VL-8B-Instruct-4bit |
defcdea7 |
Apache License 2.0 |
| rerank | mlx-community/Qwen3-VL-Reranker-2B-8bit |
e9cd8bbe |
Apache License 2.0 |
| ocr | mlx-community/Qwen3-VL-2B-Instruct-8bit |
b0338e0e |
Apache License 2.0 |
| image | mlx-community/Qwen-Image-2.1-mflux-q4 |
746a5855 |
Qwen Research License (research and non-commercial use) |
Qwen-Image-2.1 is under the Qwen Research License (research and non-commercial use), and that applies to images generated with it. Repacks of it, such as mlx-community/Qwen-Image-2.1-mflux-q4 (the image row above), are treated as Qwen Research License too, whatever their card says.
The code is licensed under the Apache License 2.0; see NOTICE. It does not cover model weights or images generated with them.