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.

Originally released by Microsoft on Hugging Face under the MIT license. 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 original repository for weights, code, and updates, and cite the paper if you build on it.

Built by

• Microsoft - Mage project team, lead author Xinjie Zhang and contributors
• Paper: Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing (arXiv:2607.19064)

Native-resolution generation

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.

Instruction-based editing

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.

Versions on Civitai

All six upstream checkpoints are mirrored here:

• Mage-Flow-4B-Base - text-to-image, 30 steps, CFG 5.0
• Mage-Flow-4B - text-to-image, RL-aligned, 20 steps, CFG 5.0
• Mage-Flow-4B-Turbo - text-to-image, 4-step distilled, CFG 1.0
• Mage-Flow-Edit-4B-Base - image editing, 30 steps, CFG 5.0
• Mage-Flow-Edit-4B - image editing, RL-aligned, 30 steps, CFG 5.0
• Mage-Flow-Edit-4B-Turbo - image editing, 4-step distilled, CFG 1.0

If you are starting out, use Mage-Flow-4B for text-to-image and Mage-Flow-Edit-4B 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.

Runs light

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.

Links

• Source: huggingface.co/microsoft/Mage-Flow-Base
• Code: github.com/microsoft/Mage
• Project page: microsoft.github.io/Mage
• Paper: arXiv:2607.19064
• License: MIT