Mage-Flow: Microsoft's 4B Native-Resolution Image Model Family
Mage-Flow is a compact 4B generative stack from Microsoft Asia for text-to-image and image editing at native resolution up to 2048, with RL-aligned and Turbo variants.
Mage-Flow by Microsoft
Text-to-ImageImage Editing4B ParametersNative ResolutionMIT LicenseMage-Flow is a compact **4B-parameter generative stack** from Microsoft Asia for efficient text-to-image generation and instruction-based image editing. Instead of scaling to tens of billions of parameters, it reaches state-of-the-art-competitive quality through careful tokenizer-backbone-system co-design, keeping it fast, memory-light, and easy to fine-tune under realistic compute budgets.
| Developer | Microsoft Asia |
| Release Date | 2026-07-22 |
| Architecture | 4B NR-MMDiT (Native-Resolution Multimodal Diffusion Transformer) + Mage-VAE |
| Text Encoder | Qwen3-VL (4B) |
| License | MIT |
| Output Resolution | 512×512 to 2048×2048 (any aspect ratio, up to 4:1) |
| Inference Steps | 4 to 30 (depends on variant) |
| VRAM | ~18-20 GB at 1024×1024 |
| GenEval Score | 0.90 (highest among open-source models) |
Variants
| Variant | Task | Steps | Description |
|---|---|---|---|
| Mage-Flow-4B | Text-to-Image | 20 | RL-aligned generation |
| Mage-Flow-4B-Turbo | Text-to-Image | 4 | Few-step distilled generation |
| Mage-Flow-Edit-4B | Image Editing | 30 | RL-aligned instruction-based editing |
| Mage-Flow-Edit-4B-Turbo | Image Editing | 4 | Few-step distilled editing |
All weights are self-contained diffusers-style checkpoints on Hugging Face. Curated ComfyUI-ready downloads (including int8 ConvRot quantized variants) are available in the Model Hub catalog.
Comments
Sign in with GitHub to join the discussion.