<p>Mage Flow is a 4-billion parameter image generation and editing stack from Microsoft. It pairs a lightweight latent tokenizer (Mage-VAE) with a native-resolution multimodal diffusion transformer (NR-MMDiT), and it generates natively from 512 to 2048 pixels at any aspect ratio - no bucket quantization, no padding. At 4B it scores 0.90 on GenEval, the best result among open-source models, against systems five to eight times its size.</p>

<p><strong>Originally released by <a href="https://huggingface.co/microsoft/Mage-Flow-Base" target="_blank" rel="noopener">Microsoft</a> on Hugging Face under the MIT license.</strong> All credit for the model goes to the Mage-Flow team. Civitai is hosting a mirror so creators can run it on-site - please head to the <a href="https://github.com/microsoft/Mage" target="_blank" rel="noopener">original repository</a> for weights, code, and updates, and cite the paper if you build on it.</p>

<h3>Built by</h3>
<ul>
  <li><a href="https://microsoft.github.io/Mage" target="_blank" rel="noopener">Microsoft - Mage project team</a>, lead author Xinjie Zhang and contributors</li>
  <li>Paper: <a href="https://arxiv.org/abs/2607.19064" target="_blank" rel="noopener">Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing</a> (arXiv:2607.19064)</li>
</ul>

<h3><strong>Native-resolution generation</strong></h3>
<p>NR-MMDiT processes variable-length image and text sequences with per-sample 2D rotary embeddings and joint self-attention. Because resolution is native rather than bucketed, extreme aspect ratios like 4:1 (512x2048, 2048x512) come out clean instead of stretched or cropped, and you are not locked to a fixed training resolution.</p>

<h3><strong>Instruction-based editing</strong></h3>
<p>The Edit variants handle semantic content changes, appearance transformation, image restoration, and structure-aware edits in one architecture - you describe the change in plain language rather than masking. On ImgEdit-Bench the RL-aligned edit model scores 4.34, with GEdit-EN at 8.127 and GEdit-CN at 8.123.</p>

<h3><strong>Versions on Civitai</strong></h3>
<p>All six upstream checkpoints are mirrored here:</p>
<ul>
  <li><strong>Mage-Flow-4B-Base</strong> - text-to-image, 30 steps, CFG 5.0</li>
  <li><strong>Mage-Flow-4B</strong> - text-to-image, RL-aligned, 20 steps, CFG 5.0</li>
  <li><strong>Mage-Flow-4B-Turbo</strong> - text-to-image, 4-step distilled, CFG 1.0</li>
  <li><strong>Mage-Flow-Edit-4B-Base</strong> - image editing, 30 steps, CFG 5.0</li>
  <li><strong>Mage-Flow-Edit-4B</strong> - image editing, RL-aligned, 30 steps, CFG 5.0</li>
  <li><strong>Mage-Flow-Edit-4B-Turbo</strong> - image editing, 4-step distilled, CFG 1.0</li>
</ul>
<p>If you are starting out, use <strong>Mage-Flow-4B</strong> for text-to-image and <strong>Mage-Flow-Edit-4B</strong> for editing. The RL-aligned builds are the quality sweet spot. Drop to the Turbo builds when you want speed and can accept a small quality tradeoff.</p>

<h3><strong>Runs light</strong></h3>
<p>Mage-VAE uses roughly 12x fewer encode and 22x fewer decode MACs per pixel than FLUX.2-VAE at matching reconstruction fidelity. Combined with native-resolution packing, which runs the conditional and unconditional CFG branches in a single packed forward pass, Mage-Flow-Turbo produces a 1024x1024 image in 0.59 seconds on a single A100 and peaks at 18-20 GB - the lowest memory footprint among the systems it was compared against.</p>

<h3>Links</h3>
<ul>
  <li><strong>Source:</strong> <a href="https://huggingface.co/microsoft/Mage-Flow-Base" target="_blank" rel="noopener">huggingface.co/microsoft/Mage-Flow-Base</a></li>
  <li><strong>Code:</strong> <a href="https://github.com/microsoft/Mage" target="_blank" rel="noopener">github.com/microsoft/Mage</a></li>
  <li><strong>Project page:</strong> <a href="https://microsoft.github.io/Mage" target="_blank" rel="noopener">microsoft.github.io/Mage</a></li>
  <li><strong>Paper:</strong> <a href="https://arxiv.org/abs/2607.19064" target="_blank" rel="noopener">arXiv:2607.19064</a></li>
  <li><strong>License:</strong> MIT</li>
</ul>

Mage Flow

Files with SHA256: 327c3967a5190ea52e453ec3dd81ba168e37a2a0ff2c763aa3e9260bbbe1913c