Microsoft has released Mage-Flow, a compact 4B-scale generative stack designed for efficient text-to-image generation and instruction-based image editing. The system achieves state-of-the-art-competitive quality through the co-design of a lightweight tokenizer, Mage-VAE, and a Native-Resolution Multimodal Diffusion Transformer (NR-MMDiT).

  • Mage-Flow matches or beats larger open systems like Qwen-Image 20B and FLUX.2 32B while using significantly fewer parameters.
  • The shared stack supports native resolution generation from 512 to 2048 pixels on any aspect ratio, including extreme 4:1 formats.
  • Mage-VAE reduces encode/decode MACs per pixel by ~12x and ~22x respectively compared to FLUX.2-VAE.
  • Fused CUDA kernels and native-resolution packing cut training time from ~1.93 s to ~0.78 s, a 2.5x speedup.
  • The release includes Base, RL-aligned, and 4-step Turbo variants for both generation (Mage-Flow) and editing (Mage-Flow-Edit).
  • On a single A100 at 1024² resolution, Mage-Flow-Turbo achieves 0.59 s/image latency with ~18–20 GB peak memory.

The architecture removes the VAE as a high-resolution bottleneck and enables interactive latency for both generation and editing tasks within realistic compute budgets.