README.md

---
license: mit
pipeline_tag: text-to-image
library_name: mage-flow
tags:
  - mirror
  - archive
  - text-to-image
  - image-generation
  - rectified-flow
  - diffusion-transformer
  - mage
---

# Mage-Flow — community mirror of `microsoft/Mage-Flow`

> **This is not my model.** This repository is an unmodified re-upload of
> **`microsoft/Mage-Flow`** (checkpoint `Mage-Flow-4B`, RL-aligned), published by the
> Microsoft Mage Team and since removed from the Hugging Face Hub. I did not train,
> fine-tune, quantize, convert or otherwise alter these weights — I am only re-hosting
> a copy I had archived, so the artifacts stay reachable for research. All credit goes
> to the original authors.
>
> I am not affiliated with Microsoft, and this mirror is not endorsed by them.
> Redistribution is done under the terms of the original MIT license (see `LICENSE`).

## Original sources

| | |
|---|---|
| Original repo | `microsoft/Mage-Flow` (no longer on the Hub) |
| Code | https://github.com/microsoft/Mage — model code under [`mage_flow/`](https://github.com/microsoft/Mage/tree/main/mage_flow) |
| Project page | https://microsoft.github.io/Mage |
| Paper | [arXiv:2607.19064](https://arxiv.org/abs/2607.19064) |
| Authors | Xinjie Zhang et al., Microsoft Mage Team |
| License | MIT (unchanged, see `LICENSE`) |

The upstream repository is the authoritative documentation — everything below is a
condensed restatement of it for orientation only.

## What this checkpoint is

Mage-Flow is a compact 4B generative stack for text-to-image generation and
instruction-based image editing, built from two parts:

- **Mage-VAE** — a symmetric one-step diffusion codec producing a 128-channel,
  16×-downsampled latent space, regularized toward FLUX.2-VAE latents.
- **Native-resolution MMDiT** — a 4B multimodal diffusion transformer trained with
  rectified flow matching, prompts encoded via Qwen3-VL, packed variable-length
  image+text sequences with per-sample 2D rotary embeddings. Native resolutions from
  512 to 2048 px at arbitrary aspect ratios, no bucket quantization.

This repo holds the **RL-aligned text-to-image variant** (`Mage-Flow-4B`, ~20 steps).
The family also included `Mage-Flow-Base` (30 steps), `Mage-Flow-Turbo` (4-step
distilled) and the three `Mage-Flow-Edit-*` editing counterparts.

## Usage

Install the upstream package from https://github.com/microsoft/Mage (directory
`mage_flow/`), then point it at this mirror instead of the original repo id:

```python
from mage_flow import MageFlowPipeline

pipe = MageFlowPipeline.from_pretrained("M-vdBerg/Mage-Flow", device="cuda")
imgs = pipe.generate(["a rain-slick street at night"], steps=20, cfg=5.0,
                     heights=[1024], widths=[1024], seeds=[42])
```

Peak memory is roughly 18–20 GB. Height and width must be multiples of 16.

## Intended use and limitations

The original authors state that these models are **released for research purposes
only and are not intended for product or service deployment**, and that use requires
appropriate human oversight in a controlled research environment. That restriction
applies to this mirror unchanged. Re-hosting adds no safety review, no evaluation and
no support on my part, and I cannot answer questions about model behaviour.

## Provenance of this copy

<!-- Keep only the lines you can actually vouch for. -->

- Downloaded from the official `microsoft/Mage-Flow` repository before its removal,
  on/around **2026-07-30**.
- Upstream revision: **<COMMIT SHA, if you have it>**
- Byte-identical to what I downloaded — no conversion, re-quantization or re-packing.
  Checksums in [`SHA256SUMS`](./SHA256SUMS).
- Microsoft's original model card is preserved verbatim as
  [`README_original.md`](./README_original.md).

## Takedown

If you hold rights to this material and want the mirror gone, open a discussion here
or contact me and I will remove it.

## Citation

Cite the original work, not this mirror:

```bibtex
@article{zhang2026mageflow,
  title={Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing},
  author={Zhang, Xinjie and others},
  journal={arXiv preprint arXiv:2607.19064},
  year={2026}
}
```