STAR: Scale-wise Text-to-image generation via Auto-Regressive representations

Xiaoxiao Ma*, Mohan Zhou*, Tao Liang, Yalong Bai, et al.

Project

• STAR is a novel scale-wise text-to-image model that is effective and efficient in performance
• Notably, STAR shows efficiency by requiring 2.95s to generate 512×512 images (compared to 6.48s for PixArt-α)

Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy

Xiaoxiao Ma, Feng Zhao, Pengyang Ling, Haibo Qiu, et al.

Project

• We revisit the sampling problem in autoregressive image generation and reveal the low and uneven information density of image tokens.
• Based on this insight, we propose an entropy-informed decoding strategy that improves both generation quality and efficiency across diverse AR models and benchmarks.

STAGE: Stable and Generalizable GRPO for Autoregressive Image Generation

Xiaoxiao Ma, Haibo Qiu, Guohui Zhang, Zhixiong Zeng, et al.

Project

• STAGE is the first system to study the stability and generalization of GRPO-based autoregressive visual generation.
• Built upon Janus-Pro-7B, STAGE improves the GenEval score from 0.78 to 0.89 (≈14%) without compromising image quality, and its effectiveness generalizes well across benchmarks such as T2I-Compbench and ImageReward.

Masked Pre-trained Model Enables Universal Zero-shot Denoiser

Xiaoxiao Ma*, Zhixiang Wei*, Yi Jin, Pengyang Ling, et al.

Project

• MPI is a zero-shot denoising pipeline designed for many types of noise degradations
• Only around 10s takes for a MPI to denoise on single noisy image

HQ-CLIP: Leveraging Large Vision-Language Models to Create High-Quality Image-Text Datasets and CLIP Models

Zhixiang Wei*, Guangting Wang*, Xiaoxiao Ma, et al.

Project

• A CLIP training framework trained on 1.3B bidirectional image–text pairs, combining bidirectional supervision and label classification, achieving SoTA zero-shot and retrieval performance.

Stronger, Fewer, \& Superior: Harnessing Vision Foundation Models for Domain Generalized Semantic Segmentation

Zhixiang Wei*, Lin Chen*, Yi Jin*, Xiaoxiao Ma, et al.

Project

• Rein is a PEFT framework based on vision foundation models for domain generalized semantic segmentation (DGSS) with merely 1% trainable parameters