---
title: "Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.19064v1"
permalink: "https://feed.astralor.com/items/itm_f02ef3d1984d1a6794d1af83"
published: "2026-07-21T12:55:34.000Z"
score: 80
channel: "ai"
tags: ["Diffusion Transformer","Mage-Flow","Mage-VAE","Native Resolution","Image Generation","Image Editing"]
---

# Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing

## AI commentary

> This paper stands out because it achieves near-SOTA high-resolution generation at 4B scale through clever co-design rather than scaling up blindly. For teams looking to deploy high-resolution image generation applications, the Turbo variant's speed and low memory footprint are directly actionable. Highly recommended for those interested in efficient AI systems, edge deployment, or interactive creative tools.

## At a glance

Mage-Flow is a compact 4B-parameter generative stack consisting of a lightweight high-fidelity latent tokenizer Mage-VAE and a native-resolution multimodal diffusion transformer. With anchor-latent regularization, native-resolution packing, and CUDA kernel fusion, it achieves ~2.5× training throughput improvement. The model family includes Base, RL-aligned, and Turbo variants; the Turbo variant generates a 1024² image in 0.59s and edits in 1.02s on a single A100, reaching competitive performance on standard benchmarks.

## Key points

- Mage-VAE uses one-step diffusion-style encoding/decoding and 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 improve end-to-end training throughput by ~2.5×.
- Diffusion-NFT enhances prompt following, text rendering, aesthetic quality, and editing fidelity; few-step distillation with adversarial perceptual guidance produces 4-step Turbo models.
- Turbo variants generate a 1024² image in 0.59s and edit in 1.02s on a single A100 with a small memory footprint.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_f02ef3d1984d1a6794d1af83