Microsoft releases Mage Flow 4B image model

// 1h agoMODEL RELEASE

Microsoft releases Mage Flow 4B image model

Microsoft has released Mage Flow, an open-source 4-billion parameter model family designed for high-efficiency text-to-image synthesis and fine-grained editing. Combining a one-step latent tokenizer (Mage-VAE) with a Native-Resolution Multimodal Diffusion Transformer (NR-MMDiT), the MIT-licensed suite supports resolutions from 512 to 2048 pixels alongside sub-second Turbo variants.

// ANALYSIS

Mage Flow proves that compact 4B models with clever architecture can rival massive 20B+ parameter image generators while drastically cutting inference costs.

–Mage-VAE enables fast single-step encoding/decoding, substantially reducing computational overhead compared to standard VAEs.
–Native resolution support from 512 to 2048 pixels handles diverse aspect ratios without relying on post-generation upscaling.
–Released under an MIT license on Hugging Face with diffusers and ComfyUI support, making it instantly accessible for community adoption and fine-tuning.

// TAGS

image-genopen-sourcediffusion-modelimage-editingmicrosofthugging-face

DISCOVERED

1h ago

2026-07-26

PUBLISHED

1h ago

2026-07-26

RELEVANCE

8/ 10

AUTHOR

AI Search

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL1h ago

Black Forest Labs previews multimodal model Flux 3

Black Forest Labs has previewed Flux 3, a unified multimodal foundation model designed to natively integrate image creation, audio synthesis, 720p video generation with up to 20 seconds of synchronized sound, and robotics action prediction. Early access features text-to-video, image-to-video, and keyframe transitions, with an open-weight community release planned.

OPEN SOURCE1h ago

Homie brings multi-view consistency to AI video

Homie is an open-source reference-to-video framework designed to solve subject and object identity drift in AI video generation. By leveraging multi-view image inputs alongside multimodal intelligent guidance, Homie maintains consistent visual features and realistic physical interactions between subjects and objects across generated video frames.

MODEL5h ago

Inflect-Micro-v2 ships standalone 9.36M parameter TTS

Inflect-Micro-v2, released by developer Owen Song on Hugging Face, is an open-weight text-to-speech (TTS) model designed for efficient local execution. Packing a complete end-to-end speech generation pipeline into 9.36 million parameters (~37.5 MB FP32), it generates 24 kHz English speech directly on CPU or CUDA hardware without external vocoders.