Mage-Flow: A 4B-parameter model for fast native-resolution image gen & edit

Mage-Flow is a compact 4B-parameter generative stack designed for efficient text-to-image generation and instruction-based image editing. It comprises two co-designed components: Mage-VAE, a lightweight high-fidelity latent tokenizer that uses one-step diffusion-style encoding/decoding with anchor-latent regularization, reducing tokenization cost by more than an order of magnitude while preserving reconstruction quality. The backbone is a Native-Resolution Multimodal Diffusion Transformer trained with rectified flow matching. Combined with native-resolution packing and stack-level CUDA kernel fusion, the system supports flexible-resolution training and improves end-to-end training throughput by about 2.5×.

The model family includes Base, RL-aligned, and Turbo variants for both generation and editing. Diffusion-NFT enhances prompt following, text rendering, aesthetic quality, and editing fidelity. Few-step distillation with adversarial perceptual guidance produces 4-step Turbo models for low-latency inference. At 1024² resolution on a single A100 GPU, Mage-Flow-Turbo generates an image in 0.59s, and Mage-Flow-Edit-Turbo edits an image in 1.02s, all with a small memory footprint. This demonstrates that careful tokenizer–backbone–system co-design can deliver strong high-resolution generation and editing at a 4B scale.