AIAny
Icon for item

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

Efficient 4B-scale image generation and editing model family that pairs a lightweight VAE tokenizer (Mage-VAE) with a native-resolution multimodal diffusion transformer, reducing tokenization cost by an order of magnitude and enabling few-step high-resolution generation and editing.

Introduction

Most high-resolution image generators scale compute and tokenization cost with pixel count, making interactive native-resolution generation expensive. Mage-Flow demonstrates that tokenizer–backbone–system co-design can push a compact 4B model family into practical native-resolution generation and editing by radically cutting tokenization cost and fusing system kernels.

Key Findings
  • Tokenizer and encoding: Mage-VAE uses one-step diffusion-style encode/decode with anchor-latent regularization, preserving strong VAE reconstruction quality while reducing tokenization cost by more than an order of magnitude — so it lowers per-image compute and memory during training and inference.
  • Training and system throughput: Native-resolution packing plus stack-level CUDA kernel fusion improve end-to-end training throughput roughly 2.5× compared with the non-co-designed baseline — so training budgets for high-res models shrink materially.
  • Model family and inference: The stack yields Base, RL-aligned, and Turbo variants. Distillation with adversarial perceptual guidance produces 4-step Turbo models; on a single NVIDIA A100 Mage-Flow-Turbo generates 1024×1024 images in ~0.59s and Mage-Flow-Edit-Turbo edits in ~1.02s — so high-res interactive latency becomes practical on one GPU.
  • Objective and training methods: A native-resolution multimodal diffusion transformer is trained with rectified flow matching; Diffusion-NFT and few-step distillation improve prompt following, text rendering, aesthetics, and editing fidelity — so smaller models can close quality gaps versus much larger generators.
What sets it apart
  • System-aware co-design: Unlike approaches that treat tokenizer, backbone, and runtime separately, Mage-Flow optimizes across all three layers to lower tokenization overhead and fuse CUDA kernels, translating model ideas into measurable throughput gains.
  • Native-resolution-first: Rather than heavily compressing to very low-dimensional latents, the stack supports flexible-resolution training with native packing, keeping high-fidelity reconstruction while retaining efficient token counts.
  • Practical low-latency variants: The Turbo 4-step distilled models are explicitly engineered for sub-second high-res generation on a single A100, targeting interactive editing/generation use cases that larger diffusion models usually cannot meet without multiple GPUs.
Who it's for — tradeoffs and limits

Great fit if you need interactive high-resolution image generation or editing on single-GPU hardware, want a compact (≈4B) model family with few-step inference, or value system-level efficiency improvements that lower training and inference costs. Look elsewhere if absolute state-of-the-art unconstrained FID/IS on massive compute is your sole priority: a compact 4B family will still trade some headroom versus very large (tens of billions) models trained without such compute constraints. Also, organizations requiring audited or well-documented author affiliations may need to consult the paper or its code release for full reproducibility and institutional metadata.

Where it fits

Mage-Flow positions as a practical, efficiency-focused alternative to large-scale native-resolution diffusion stacks and to standard latent-diffusion pipelines: it narrows the gap between compact foundation models and high-resolution workflows by emphasizing tokenizer/backbone/system co-optimization rather than only scaling parameter counts.

Information

  • Websitearxiv.org
  • AuthorsXinjie Zhang, Peng Zhang, Shicheng Zheng, Jinghao Guo, Zhaoyang Jia, Yifei Shen, Xun Guo, Yuxuan Luo, Jiahao Li, Wenxuan Xie
  • Published date2026/07/21

More Items

Analyzes internal computation of text-to-image diffusion transformers and shows structural template tokens act as implicit semantic registers that maintain object identity during denoising. Introduces a causal interpretability framework (attention decomposition + targeted interventions) and a training-free pruning rule that cuts ~20% attention FLOPs for a ~1.4-point GenEval drop.

Synthesizes RGB frames from structured world states exported by physics engines; it reformulates a heavy generative renderer into a few-step autoregressive streaming model and uses lightweight distilled codecs to reach playable ~30 FPS while preserving G-buffer and prompt control.

Personalizes subject-driven videos to preserve human identity and accurate human–object interactions by integrating multimodal references and MLLM-derived semantics. Introduces global multimodal guidance in self-attention and modality-reference embeddings to align MLLM features with VAE tokens, supporting both inter- and intra-subject inputs (e.g., OCR, multi-view).