SRC-Flow: Compact Semantic Representations Enable Normalizing Flows for Image Generation

Longtao Jiang, Jianmin Bao, Zhendong Wang, Xin Tao, Zhihui Li, Xiaojun Chang

arXiv Under review

• We identify a semantic-capacity mismatch between normalizing flows and high-dimensional pretrained vision features, and propose a Semantic Representation Compressor that compresses frozen RAE features into a compact 32-channel space for flow modeling.
• With constant noise regularization, SRC-Flow achieves state-of-the-art gFID among normalizing flows on ImageNet 256×256 (1.65) and 512×512 (2.07).

Token Painter: Training-Free Text-Guided Image Inpainting via Mask Autoregressive Models

Longtao Jiang, Jie Huang, Mingfei Han, Lei Chen, Yongqiang Yu, Feng Zhao, Xiaojun Chang, Zhihui Li

arXiv Paper Code

• We propose Token Painter, a training-free text-guided image inpainting method built on Mask AutoRegressive (MAR) models.
• A Dual-Stream Encoder Information Fusion (DEIF) module fuses text and background information in the frequency domain to produce guidance tokens, and an Adaptive Decoder Attention Score Enhancing (ADAE) module strengthens attention on guidance and inpainting tokens.

SmartEraser: Remove Anything from Images using Masked-Region Guidance

Longtao Jiang, Zhendong Wang, Jianmin Bao, Wengang Zhou, Dong Chen, Houqiang Li

arXiv Paper Code

• We introduce SmartEraser, built with a new “removing” paradigm called Masked-Region Guidance.
• The paradigm guides the model to accurately identify the object to be removed, preventing its regeneration in the output.
• Since the user mask often extends beyond the object itself, it aids in preserving the surrounding context in the final result.

Structure-guided Diffusion Transformer for Low-Light Image Enhancement

Xiangchen Yin, Zhenda Yu, Longtao Jiang, Xin Gao, Xiao Sun, Zhi Liu, Xun Yang

arXiv Paper

• We introduce DiT into the low-light enhancement task and design a novel Structure-guided Diffusion Transformer based Low-light image enhancement (SDTL) framework.
• Extensive qualitative and quantitative experiments demonstrate that our method achieves SOTA performance on several popular datasets.

SEDS: Semantically Enhanced Dual-Stream Encoder for Sign Language Retrieval

Longtao Jiang, Min Wang, Zecheng Li, Yao Fang, Wengang Zhou, Houqiang Li

arXiv Paper Code

• We propose a novel representation framework called Semantically Enhanced Dual-Stream Encoder (SEDS), which aggregates Pose and RGB modalities to represent local and global information of sign language videos.
• The aggregated clip-level video features are fed into the CLIP vision encoder for interaction, and matched with the word-level text features embedded using the CLIP text encoder.