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/
Project page	https://microsoft.github.io/Mage
Paper	arXiv: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:

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

• 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.
• Microsoft's original model card is preserved verbatim as 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:

@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}
}