MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation

Yawen Shao; Jie Xiao; Kai Zhu; Yu Liu; Hongchen Luo; Xueyang Fu; Yang Cao; Wei Zhai

cs.CVOct 12, 2026

MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation

Authors: Yawen Shao, Jie Xiao, Kai Zhu, Yu Liu, Hongchen Luo, Xueyang Fu, Yang Cao, Wei Zhai,

Organizations: University of Science and Technology of China · Tongyi Lab, Alibaba Group · Northeastern University

Abstract

Reinforcement learning (RL) has shown strong potential for aligning generative models with human intent, but adapting RL to masked generative models (MGMs) remains largely underexplored. Existing approaches that directly adapt Group Relative Policy Optimization (GRPO) to MGMs leave a foundational conflict unaddressed: MGMs' exploitative decoding nature is inherently at odds with the exploratory diversity that GRPO presupposes. In this work, we reveal a structural decoding bias termed the visual exploration trap: confidence-based sampling trades the exploration of diverse subject realizations for the greedy resolution of low-uncertainty background regions, causing premature collapse of the generative exploration space and starving GRPO of the rollout diversity it depends on. To this end, we propose MaskSense, a novel RL framework for MGMs that confronts the visual exploration trap at both the sampling and optimization stages. Specifically, we introduce a semantic anchored routing sampling mechanism that leverages semantic priors to preserve high-entropy subject token exploration while amplifying intra-group reward variance through exploratory-exploitative routing to yield more discriminative advantage estimates. Furthermore, we design a global entropy transition anchoring strategy to identify the most consequential decoding steps and concentrate gradient updates on them. Extensive experiments on multiple text-to-image benchmarks demonstrate that MaskSense substantially improves upon the base model, with GenEval accuracy improving from 54% to 86% and HPS from 28.89 to 36.39, achieving state-of-the-art performance.

Figures & tables

Figure 1: Motivation. (a) MaskGIT keeps subject tokens masked while committing background regions early, whereas SA-Sampler resolves subject regions earlier via semantic anchoring. (b) SA-Sampler consistently discovers higher-reward trajectories through broader generative exploration. (c) SA-Sampler maintains higher subject-token fixation ratios and elevated entropy, effectively escaping the visual exploration trap.

Figure 2: Overview of MaskSense. MaskSense comprises two core modules that jointly address the visual exploration trap. At the sampling stage, SARS promotes subject-token exploration via semantic anchored decoding and sharpens advantage estimation through dual-branch rollout routing. At the optimization stage, GETA concentrates GRPO updates on critical decoding steps anchored by entropy transitions, with random remasking applied to ensure optimization diversity.

Diffusion-based Models
Model	Overall $\uparrow$	Single Obj. $\uparrow$	Two Obj. $\uparrow$	Counting $\uparrow$	Color $\uparrow$	Position $\uparrow$	Color Attr. $\uparrow$
LDM [ 38 ]	0.37	0.92	0.29	0.23	0.70	0.02	0.05
SD-XL [ 34 ]	0.55	0.98	0.74	0.39	0.85	0.15	0.23
SD3.5-M [ 10 ]	0.63	0.98	0.78	0.50	0.81	0.24	0.52
DALLE-3 [ 3 ]	0.67	0.96	0.87	0.47	0.83	0.43	0.45
FLUX.1-dev [ 22 ]	0.66	0.98	0.81	0.74	0.79	0.22	0.45

Table 1: Quantitative Comparison on the GenEval. Best scores are bolded . Results for other models are from [ 69 ] .

\Delta

indicates improvements over Meissonic. Obj.: Object; Attr.: Attribution.

Panel of Figure 3 — Figure 3: Comparison of Training Dynamics Curves. Fig.(a) to (c) refer to the tasks of compositional image generation, human preference alignment, and multi-reward optimization, respectively.

Figure 4: Qualitative Comparison. MaskSense achieves superior visual quality with richer fine-grained details and more faithful satisfaction of object-level constraints across both tasks.

Table 6

Table 4: Hyper-parameter Study . We investigate the influence of the semantic strength

\lambda

, the temporal gate

\tau

in the SARS, and the number of anchor steps

k

in the GETA, respectively.

Panel of Figure 5 — Figure 5: Ablation Analysis on Key Strategy Choices. (a) SARS effectively balances exploration and exploitation, outperforming single-branch sampling strategies. (b) GETA anchors critical decision steps, enabling more principled policy optimization than heuristic alternatives. (c) Random remasking diversifies anchor states, producing more stable optimization.

Figure 6: Anchored Timesteps of Diverse Prompts. (a) Global entropy and absolute entropy difference between consecutive timesteps. (b) Visualization of generation process across timesteps.

Step	0	200	400	600	800	1K	1.2K	1.4K	1.6K	1.8K	2K
Mask-GRPO HPS	26.54	27.28	28.09	28.54	29.05	29.44	29.56	29.72	29.74	29.72	29.81
MaskSense HPS	26.54	27.60	28.22	28.92	29.30	29.66	29.82	29.98	29.97	30.05	30.09
Mask-GRPO GPU Hours	0.00	11.10	22.17	33.21	44.30	55.37	66.44	77.51	88.61	99.66	110.72
MaskSense GPU Hours	0.00	11.45	22.89	34.31	45.73	57.15	68.56	79.96	91.37	102.77	114.18

Table 5: Transfer to Show-o on HPS. Evaluation HPS scores and cumulative training GPU hours.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 7 — Figure 7: Comparison of Compute Overhead Dynamics. Fig.(a) to (c) refer to the tasks of compositional image generation, human preference alignment, and multi-reward optimization, respectively.

Figure 8: Sampling Process Visualization of MaskGIT Sampler on Meissonic baseline.

Figure 9: Sampling Process Visualization of SA-Sampler on Meissonic baseline.

Figure 10: Sampling Process Visualization of MaskGIT Sampler on MaskSense after RL Training.

Figure 11: Qualitative Comparisons on Human Preference Alignment. MaskSense generates images with richer stylistic details, superior texture fidelity, and more faithful rendering of complex artistic styles compared to Meissonic and GRPO baselines.

Figure 12: Qualitative Comparisons on Compositional Image Generation. MaskSense more reliably satisfies compositional constraints including object count, spatial arrangement, and color attribution across GenEval and GenEval+CLIP settings.

Generative Modeling

Generative modeling learns data distributions to create new samples, from images and text to molecules, simulations, and structured records.

See all 500 papers

Accelerating Masked Image Generation by Learning Controlled Latent Dynamics
Feb 27, 2026Kaiwen Zhu, Quansheng Zeng, Yuandong Pu +9Latent Dynamics Modeling T2I Generation
Shanghai Jiao Tong University Shanghai, China · Shanghai AI Laboratory Shanghai, China · Shanghai Innovation Institute Shanghai, China +6
Fixed-Point Masked Generative Modeling
May 29, 2026Andrea Miele, Yiming Qin, Alba Carballo-Castro +2
LTS4, EPFL · CLAIRE, EPFL
Mask-Aware Policy Gradients for Diffusion Language Models
Jul 16, 2026Haran Raajesh, Kulin Shah, Adam Klivans +1Masked Diffusion Models RL for Language Model Reasoning
Department of Computer Science The University of Texas at Austin

Feb 27, 2026J/K move · Enter open · S save

Accelerating Masked Image Generation by Learning Controlled Latent Dynamics

Kaiwen Zhu, Quansheng Zeng, Yuandong Pu, Shuo Cao+8

Shanghai Jiao Tong University Shanghai, China · Shanghai AI Laboratory Shanghai, China · Shanghai Innovation Institute Shanghai, China+6

Latent Dynamics Modeling T2I Generation Masked Image Modeling

Open paper AskPDF arXiv

Masked Image Generation Models (MIGMs) have achieved great success, yet their efficiency is hampered by the multiple steps of bi-directional attention. In fact, there exists notable redundancy in their computation: when sampling discrete tokens, the rich semantics contained in the continuous features are lost. Some existing works attempt to cache the features to approximate future features. However, they exhibit considerable approximation error under aggressive acceleration settings. We attribute this to their limited expressivity and the failure to account for sampling information. To fill this gap, we propose learning a lightweight model that incorporates both previous features and sampled tokens, and regresses the average velocity field of feature evolution. The model has moderate complexity that suffices to capture the subtle dynamics while keeping lightweight compared to the original base model. We apply our method to two representative MIGMs and tasks. In particular, on the state-of-the-art Lumina-DiMOO, it achieves over 4x acceleration of text-to-image generation while maintaining quality, significantly pushing the Pareto frontier of masked image generation. The code and model weights are available at https://github.com/Kaiwen-Zhu/MIGM-Shortcut.

Abstract

Generative Modeling

Explore similar work

Accelerating Masked Image Generation by Learning Controlled Latent Dynamics

Fixed-Point Masked Generative Modeling

Mask-Aware Policy Gradients for Diffusion Language Models