Mage-Flow: Efficient Native-Resolution Foundation Model for Image Generation
4 hours ago
- Mage-Flow is a compact 4B-scale generative stack for efficient text-to-image generation and instruction-based image editing.
- It consists of two co-designed components: Mage-VAE (a lightweight high-fidelity latent tokenizer) and a Native-Resolution Multimodal Diffusion Transformer.
- Mage-VAE uses one-step diffusion-style encoding/decoding with anchor-latent regularization, reducing tokenization cost by over an order of magnitude while preserving reconstruction quality.
- Native-resolution packing and stack-level CUDA kernel fusion enable flexible resolution training and improve end-to-end throughput by about 2.5×.
- The model family includes Base, RL-aligned, and Turbo variants for both generation and editing, with 4-step Turbo models achieving low-latency inference (0.59s for generation, 1.02s for editing at 1024² on a single A100).
- Mage-Flow-Turbo achieves competitive benchmark scores (GenEval 0.88, CVTG-2K 0.873) despite its compact size, outperforming larger models like Qwen-Image (20B) and FLUX.2-Klein-9B.