Mask-Aware Flow Matching for Zero-Destruction Text Editing
Implement localized Region of Interest mask-modulated cross-attention and velocity penalty losses in MMDiT architectures to prevent background text destruction during image editing.
Source paper
Changyao Tian, Danni Yang, Guanzhou Chen, Erfei Cui, Zhaokai Wang, Yuchen Duan, Penghao Yin, Sitao Chen, Ganlin Yang, Mingxin Liu, Zirun Zhu, Ziqian Fan, Leyao Gu, Haomin Wang, Qi Wei, Jinhui Yin, Xue Yang, Zhihang Zhong, Qi Qin, Yi Xin, Bin Fu, Yihao Liu, Jiaye Ge, Qipeng Guo, Gen Luo, Hongsheng Li, Yu Qiao, Kai Chen, Hongjie Zhang
Generated by AI. The paper’s authors did not write these.
Implement localized Region of Interest mask-modulated cross-attention and velocity penalty losses in MMDiT architectures to prevent background text destruction during image editing.
Replace static variance normalization with a token-adaptive gating projection network to selectively manage outlier magnitudes before visual generation heads.
Generate synthetic physics image editing pairs using scriptable 3D simulation engines and fine-tune an image-to-image translation adapter to bridge the sim-to-real gap.
This paper introduces InternVL-U, a lightweight 4B-parameter unified multimodal model designed to integrate visual understanding, reasoning, generation, and editing. By combining an open-source MLLM backbone with a custom continuous visual generation head and leveraging a reasoning-centric Chain-of-Thought (CoT) data synthesis pipeline, the model bridges the gap between aesthetic image generation and high-level semantic intelligence.
The architecture features unified contextual modeling using an encoder-based MLLM initialized with ViT features, paired with a specialized Dual-Stream MMDiT visual generation head that predicts VAE latents via Flow Matching. It employs an asymmetric, decoupled visual representation strategy and a multi-stage training curriculum to balance autoregressive text generation with continuous image synthesis.