Discrete Diffusion Language Models have emerged as a compelling paradigm for unified multimodal generation, yet their deployment is hindered by high inference latency arising from iterative decoding. Existing acceleration strategies often require expensive re-training or fail to leverage the 2D spatial redundancy inherent in visual data. To address this, we propose Locality-Aware Dynamic Rescue (LADR), a training-free method that expedites inference by exploiting the spatial Markov property of images. LADR prioritizes the recovery of tokens at the ''generation frontier'', regions spatially adjacent to observed pixels, thereby maximizing information gain. Specifically, our method integrates morphological neighbor identification to locate candidate tokens, employs a risk-bounded filtering mechanism to prevent error propagation, and utilizes manifold-consistent inverse scheduling to align the diffusion trajectory with the accelerated mask density. Extensive experiments on four text-to-image generation benchmarks demonstrate that our LADR achieves an approximate 4 x speedup over standard baselines. Remarkably, it maintains or even enhances generative fidelity, particularly in spatial reasoning tasks, offering a state-of-the-art trade-off between efficiency and quality.
Figures & tables
Figure 1: Comparison between Standard Parallel Decoding and our LADR method. While standard parallel decoding follows a fixed schedule, LADR accelerates decoding by exploiting spatial locality to dynamically recover neighbor tokens and keeps generation quality.
Figure 2: Overview of the LADR method. At each timestep, the flattened discrete tokens were reshaped into a 2D grid to identify candidate neighbors adjacent to resolved regions. These candidates are evaluated using the Confidence Margin (confidence top1-top2 gap) and are dynamically "rescued" (unmasked) based on an adaptive rescue ratio α and threshold γ . To synchronize the generation timeline with this accelerated accumulation of tokens, the Trajectory Re-alignment module utilizes an inverse cosine function to re-calculate the effective timestep t2 , allowing the scheduler to skip redundant iterations while maintaining consistency.
Figure 3: Visualization of localized semantic changes caused by perturbing a small set of VQ tokens. The impact remains spatially confined, supporting the locality assumption that nearby tokens dominate information gain.
Method
Avg. t (s) ↓
Two Obj.
Colors
Attribute
Single Obj.
Position
Counting
Overall ↑
Lumina-DiMOO
57.01
93.94
91.49
73.00
97.50
79.00
85.00
86.66
+ ML-Cache
31.95
93.75
89.63
75.50
100.00
84.50
85.94
87.83
+ Prophet
32.15
91.16
86.44
70.75
96.56
74.50
84.06
83.91
+ LADR(Ours)
13.22
91.41
91.22
74.75
99.06
85.50
81.88
87.30
Table 1: Performance comparison on the GenEval Ghosh et al. (2023) .
Method
Avg. t (s) ↓
Style
Know.
Attr.
Action
Rel.
Cmp.
Gram.
Logic.
Lay.
Text
Lumina-DiMOO
57.21
91.52
89.87
79.29
71.48
78.55
73.45
69.79
43.58
85.63
27.87
+ ML-Cache
31.96
91.40
88.77
79.17
73.67
79.06
74.61
69.65
45.87
83.96
26.15
+ Prophet
32.47
87.40
84.34
75.64
68.25
75.63
65.85
64.97
39.91
83.02
25.57
+ LADR( Ours )
13.94
94.80
88.61
81.73
75.95
81.98
77.71
66.31
45.41
85.26
16.38
Table 2: Performance comparison on UniGen-Bench Wang et al. (2025d) .
Method
Avg. t (s) ↓
Color
Shape
Texture
Spatial
Non-spatial
Complex
Lumina-DiMOO
56.92
81.07
57.02
69.30
46.70
31.70
34.98
+ ML-Cache
31.78
81.52
57.75
70.28
46.79
31.83
35.23
+ Prophet
32.43
80.92
55.48
70.34
42.37
31.46
34.92
+ LADR( Ours )
13.41
82.25
58.94
72.34
46.35
31.60
36.19
Table 3: Performance Comparison on T2I-CompBench Huang et al. (2023) .
Method
Avg. t (s) ↓
Global
Entity
Attribute
Relation
Other
Overall ↑
Lumina-DiMOO
58.16
77.20
90.36
87.93
93.04
82.80
83.61
+ ML-Cache
32.01
81.46
90.37
88.16
93.27
84.40
84.02
+ Prophet
34.63
81.76
89.56
87.33
92.76
83.20
82.91
+ LADR( Ours )
14.52
84.19
91.47
89.12
94.20
81.20
85.42
Table 4: Performance evaluation on DPG-Bench Hu et al. (2024) .
Method
Avg.t (s) ↓
Text
Lumina-DiMOO
57.21
27.87
LADR
13.94
16.38
LADR(delayed rescue)
22.51
26.10
Table 5: Efficiency and performance comparison of different strategies on the text category of UniGen-Bench.
Figure 4: Qualitative comparison of different methods in terms of generation fidelity and inference time, where the corresponding text prompt is provided below each row, with the inference latency displayed in parentheses under the image.
Figure 5: Ablation study of spatial selection strategies on the GenEval benchmark.“Random” means Random Selection , “Non-neighbor” represents Non-Neighbor Prioritization , and “Neighbor (Ours)” is our strategy that rescues neighbor tokens.
Figure 6: Ablation study about different ranking metrics of neighbor tokens on the GenEval benchmark.“Random” means Random Neighbor Selection , “Confidence” denotes Standard Confidence , and “Top1-Top2 (Ours)” means confidence margin we employed.
Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.
Existing text-to-image diffusion models excel at generating high-quality images, but face significant efficiency challenges when scaled to high resolutions, like 4K image generation. While previous research accelerates diffusion models in various aspects, it seldom handles the inherent redundancy within the latent space. To bridge this gap, this paper introduces DC-Gen, a general framework that accelerates text-to-image diffusion models by leveraging a deeply compressed latent space. Rather than a costly training-from-scratch approach, DC-Gen uses an efficient post-training pipeline to preserve the quality of the base model. A key challenge in this paradigm is the representation gap between the base model's latent space and a deeply compressed latent space, which can lead to instability during direct fine-tuning. To overcome this, DC-Gen first bridges the representation gap with a lightweight embedding alignment training. Once the latent embeddings are aligned, only a small amount of LoRA fine-tuning is needed to unlock the base model's inherent generation quality. We verify DC-Gen's effectiveness on SANA and FLUX.1-Krea. The resulting DC-Gen-SANA and DC-Gen-FLUX models achieve quality comparable to their base models but with a significant speedup. Specifically, DC-Gen-FLUX reduces the latency of 4K image generation by 53x on the NVIDIA H100 GPU. When combined with NVFP4 SVDQuant, DC-Gen-FLUX generates a 4K image in just 3.5 seconds on a single NVIDIA 5090 GPU, achieving a total latency reduction of 138x compared to the base FLUX.1-Krea model. Code: https://github.com/dc-ai-projects/DC-Gen.
Masked Discrete Diffusion Models (MDMs) have achieved strong performance across a wide range of multimodal tasks, including image understanding, generation, and editing. However, their inference speed remains suboptimal due to the need to repeatedly process redundant masked tokens at every sampling step. In this work, we propose Sparse-LaViDa, a novel modeling framework that dynamically truncates unnecessary masked tokens at each inference step to accelerate MDM sampling. To preserve generation quality, we introduce specialized register tokens that serve as compact representations for the truncated tokens. Furthermore, to ensure consistency between training and inference, we design a specialized attention mask that faithfully matches the truncated sampling procedure during training. Built upon the state-of-the-art unified MDM LaViDa-O, Sparse-LaViDa achieves up to a 2x speedup across diverse tasks including text-to-image generation, image editing, and mathematical reasoning, while maintaining generation quality.
Diffusion Transformers (DiTs) are a dominant backbone for high-fidelity text-to-image generation due to strong scalability and alignment at high resolutions. However, quadratic self-attention over dense spatial tokens leads to high inference latency and limits deployment. We observe that denoising is spatially non-uniform with respect to aesthetic descriptors in the prompt. Regions associated with aesthetic tokens receive concentrated cross-attention and show larger temporal variation, while low-affinity regions evolve smoothly with redundant computation. Based on this insight, we propose AccelAes, a training-free framework that accelerates DiTs through aesthetics-aware spatio-temporal reduction while improving perceptual aesthetics. AccelAes builds AesMask, a one-shot aesthetic focus mask derived from prompt semantics and cross-attention signals. When localized computation is feasible, SkipSparse reallocates computation and guidance to masked regions. We further reduce temporal redundancy using a lightweight step-level prediction cache that periodically replaces full Transformer evaluations. Experiments on representative DiT families show consistent acceleration and improved aesthetics-oriented quality. On Lumina-Next, AccelAes achieves a 2.11× speedup and improves ImageReward by +11.9% over the dense baseline. Code is available at https://github.com/xuanhuayin/AccelAes.
Xuanhua Yin, Chuanzhi Xu, Haoxian Zhou +2
School of Computer Science, The University of Sydney, NSW 2006, Australia