Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing
Mage-Flow claims interactive-speed 1024² image generation and editing from a compact 4B model family.
The paper pairs a lightweight Mage-VAE tokenizer with a native-resolution multimodal diffusion transformer trained using rectified flow matching. It says the VAE cuts tokenization cost by more than an order of magnitude while preserving reconstruction quality, and the full stack improves training throughput by about 2.5x. The authors report Base, RL-aligned, and Turbo variants for generation and instruction-based editing. On a single NVIDIA A100, the Turbo model generates a 1024² image in 0.59 seconds, while the edit model completes in 1.02 seconds. HF Daily Papers' note
The paper pairs a lightweight Mage-VAE tokenizer with a native-resolution multimodal diffusion transformer trained using rectified flow matching. It says the VAE cuts tokenization cost by more than an order of magnitude while preserving reconstruction quality, and the full stack improves training throughput by about 2.5x. The authors report Base, RL-aligned, and Turbo variants for generation and instruction-based editing. On a single NVIDIA A100, the Turbo model generates a 1024² image in 0.59 seconds, while the edit model completes in 1.02 seconds. HF Daily Papers' note
score 5