Mini toolsMage Flow
By Microsoft
Mage-Flow is a 4B-parameter foundation model for text-to-image generation, part of a family that also includes Base and Turbo variants. It pairs Mage-VAE, a lightweight one-step diffusion latent tokenizer, with NR-MMDiT, a native-resolution multimodal diffusion transformer trained with rectified flow matching. Native-resolution packing lets one checkpoint generate images from 512 to 2048 pixels at any aspect ratio, including extreme ratios like 4:1, without bucket quantization or padding. This RL-aligned version is optimized with Diffusion-NFT for improved prompt following, aesthetics, and text rendering, scoring competitively against larger open models such as Qwen-Image (20B) and FLUX.2 (32B) on benchmarks like GenEval and CVTG-2K.
Overview
Mage-Flow is a compact 4B-parameter text-to-image diffusion model from Microsoft, built on a Native-Resolution Multimodal Diffusion Transformer (NR-MMDiT) paired with the Mage-VAE tokenizer. It generates images natively from 512 to 2048 pixels at any aspect ratio using rectified flow matching, reaching quality competitive with much larger open models while running efficiently.
About Microsoft
Microsoft is a technology company that offers a wide range of software, cloud computing services, hardware, and artificial intelligence solutions.
Industry:
Technology, Information and Internet
Company Size:
228000
Location:
Redmond, Washington, US
Website:
microsoft.com
Last updated: July 22, 2026
