Search papers, labs, and topics across Lattice.
This paper introduces Mage-Flow, a compact 4B-scale generative model designed for efficient text-to-image generation and instruction-based image editing, addressing the high costs associated with large-scale visual generators. The model integrates Mage-VAE, a lightweight latent tokenizer that significantly reduces tokenization costs while preserving reconstruction quality, and a Native-Resolution Multimodal Diffusion Transformer that enhances training throughput by approximately 2.5 times. Mage-Flow's Turbo variants enable high-resolution generation and editing at interactive speeds, achieving competitive performance on standard benchmarks with a minimal memory footprint.
Mage-Flow achieves high-resolution image generation and editing in under a second on a single GPU, challenging the notion that larger models are always necessary for quality.
Large-scale visual generators are increasingly capable but costly to train, fine-tune, and deploy. We introduce Mage-Flow, a compact 4B-scale generative stack for efficient text-to-image generation and instruction-based image editing. The stack is built from two co-designed components: Mage-VAE, a lightweight high-fidelity latent tokenizer, and a Native-Resolution Multimodal Diffusion Transformer trained with rectified flow matching. Mage-VAE uses one-step diffusion-style encoding and decoding with anchor-latent regularization, preserving the reconstruction quality of strong public VAEs while reducing tokenization cost by more than an order of magnitude. Together with native-resolution packing and stack-level CUDA kernel fusion, the stack supports flexible-resolution training and improves end-to-end training throughput by about $2.5\times$. Built on this foundation, we develop a complete model family with Base, RL-aligned, and Turbo variants for both generation and editing. Diffusion-NFT improves prompt following, text rendering, aesthetic quality, and editing fidelity, while few-step distillation with adversarial perceptual guidance produces 4-step Turbo models for low-latency inference. Despite its compact scale, Mage-Flow and Mage-Flow-Edit achieves competitive performance across standard generation and editing benchmarks. More importantly, the Turbo variants make high-resolution generation and editing practical for interactive use: at $1024^2$ resolution on a single NVIDIA A100 GPU, Mage-Flow-Turbo generates an image in 0.59s, and Mage-Flow-Edit-Turbo edits an image in 1.02s, while maintaining a small memory footprint. These results show that careful tokenizer--backbone--system co-design can deliver strong high-resolution generation and editing within an efficient 4B model family.
Microsoft Research