Organizations: Tsinghua University · Shanghai Jiao Tong University · Shanghai AI Laboratory · Sichuan University · The Hong Kong University of Science and Technology (Guangzhou) · University of Chicago
Native unified multimodal models, which integrate both generative and understanding capabilities, face substantial computational overhead that hinders their real-world deployment. Existing acceleration techniques typically employ a static, monolithic strategy, ignoring the fundamental divergence in computational profiles between iterative generation tasks (e.g., image generation) and single-pass understanding tasks (e.g., VQA). In this work, we present the first systematic analysis of unified models, revealing pronounced parameter specialization, where distinct neuron sets are critical for each task. This implies that, at the parameter level, unified models have implicitly internalized separate inference pathways for generation and understanding within a single architecture. Based on these insights, we introduce a training-free and task-aware acceleration framework, FlashU, that tailors optimization to each task's demands. Across both tasks, we introduce Task-Specific Network Pruning and Dynamic Layer Skipping, aiming to eliminate inter-layer and task-specific redundancy. For visual generation, we implement a time-varying control signal for the guidance scale and a temporal approximation for the diffusion head via Diffusion Head Cache. For multimodal understanding, building upon the pruned model, we introduce Dynamic Token Pruning via a V-Norm Proxy to exploit the spatial redundancy of visual inputs. Extensive experiments on Show-o2 demonstrate that FlashU achieves 1.78× to 2.01× inference acceleration across both understanding and generation tasks while maintaining SOTA performance, outperforming competing unified models and validating our task-aware acceleration paradigm. Our code is publicly available at https://github.com/Rirayh/FlashU.
Figures & tables
Figure 1: Task-specific FFN neuron analysis for generation and understanding. Each layer shows the fraction of neurons that are specific to generation, understanding, or shared between both tasks (excluding unimportant neurons). Percentages indicate the proportion of neurons in each layer ranked among the top 50% for task importance, as measured by OBD-inspired sensitivity scores Δi .
Figure 2: Understanding Redundancy. Left: Intra-layer cosine similarity between the input and output features of each layer for newly generated tokens in understanding tasks. Right: Inter-layer heatmap of the cosine similarity among layer outputs for newly generated tokens. All measurements were conducted on samples from the MME [ 19 ] .
Figure 3: Illustration of the FlashU acceleration framework. (a) Task-Specific FFN Pruning calculates an importance score Ij based on activation norms and weight magnitude to statically mask redundant neurons for a specific task. (b) Dynamic Token Pruning utilizes the V-Norm proxy at shallow layers (the 2nd layer) to identify and prune visually spatially redundant tokens. (c) Dynamic Layer Skipping bypasses layers exhibiting high input-output cosine similarity ( Si ), re-evaluating the skipping mask every TLS steps. (d) Diffusion Head Cache exploits temporal coherence in the generative process, computing the full diffusion head only at fixed intervals and reusing cached hidden states for intermediate steps.
Figure 4: Redundancy in Generation Tasks. Left: Intra-layer cosine similarity between the input and output of each layer. Right: Inter-layer heatmap of hidden state cosine similarity. All measurements were conducted on samples from the GenEval [ 21 ] .
Figure 5: The relation between attention score and the norm of the value matrix for each token. Tokens with the largest attention score tend to have a lower norm in the value matrix.
Models
# Params.
MME ↑
GQA ↑
MMB ↑
MMMU ↑
MMStar ↑
AI2D ↑
Latency (s) ↓
(p)
(en)
(val)
MetaMorph [ 58 ]
8B
-
-
75.2
-
-
-
-
ILLUME [ 60 ]
7B
1445.3
-
75.1
38.2
-
71.4
-
Show-o [ 71 ]
1.3B
1097.2
58.0
-
27.4
-
-
-
JanusFlow [ 46 ]
1.5B
1333.1
60.3
74.9
29.3
-
-
-
SynerGen-VL [ 33 ]
2.4B
1381.0
-
53.7
34.2
-
-
-
Table 1: Evaluation on multimodal understanding benchmarks. # Params. indicates the number of parameters of the base LLM. Results in gray indicate the vanilla Show-o2 performance. Latency is measured in seconds (s), conducted on samples from MMMU.
Method
# Params.
Single Obj.
Two Obj.
Counting
Colors
Position
Color Attri.
Overall ↑
Latency (s) ↓
SEED-X [ 20 ]
17B
0.97
0.58
0.26
0.80
0.19
0.14
0.49
-
TokenFlow-XL [ 48 ]
14B
0.95
0.60
0.41
0.81
0.16
0.24
0.55
-
ILLUME [ 60 ]
7B
0.99
0.86
0.45
0.71
0.39
0.28
0.61
-
Show-o [ 71 ]
1.3B
0.98
0.80
0.66
0.84
0.31
0.50
0.68
-
MUSE-VL [ 73 ]
7B
-
-
-
-
-
-
0.57
-
Transfusion [ 83 ]
3.5B
-
-
-
-
-
-
0.63
-
Table 2: Evaluation on the GenEval [ 21 ] . # Params. indicates the number of parameters of the base LLM. Results in gray indicate the vanilla Show-o2 performance. The subscript values show the speedup achieved by FlashU. Latency is measured in seconds (s).
Figure 6: Comparison of four unified multimodal models on generating a three-panel canvas with seamless cross-panel continuity. While baseline models fail to satisfy the structural or continuity requirements, our FlashU 1.5B (from Show-o2 [ 72 ] ) accurately produces all three aligned segments with the fastest inference time.
Method
# Params.
Global
Entity
Attribute
Relation
Overall ↑
Latency (s) ↓
Hunyuan-DiT [ 34 ]
1.5B
84.59
80.59
88.01
74.36
78.87
-
Playground v2.5 [ 32 ]
-
83.06
82.59
81.20
84.08
75.47
-
PixArt- Σ [ 5 ]
-
86.89
82.89
88.94
86.59
80.54
-
DALL-E 3 [ 2 ]
-
90.97
89.61
88.39
90.58
83.50
-
SD3-Medium [ 18 ]
2B
87.90
91.01
88.83
80.70
84.08
-
Emu3-DPO [ 62 ]
8B
-
-
-
-
81.60
124.8
Table 3: Evaluation on DPG-Bench [ 23 ] . # Params. denotes the parameter count of the base LLM. Results in gray correspond to vanilla Show-o2. Subscripts indicate the speedup of FlashU. Latency is reported in seconds (s), averaged per sample.
Strategy
Latency (s)
Score
Show-o2 [ 72 ]
11.30
86.14
FlashU (Ours)
5.89 +1.92 ×
84.39
w/o Dynamic Layer Pruning
6.70 +1.69 ×
85.56
w/o Hybrid Network
5.65 +2.00 ×
83.68
w/o Adaptive Guidance
8.86 +1.28 ×
85.34
FFN Pruning Only
9.21 +1.23 ×
85.08
Table 4: Ablation study on the 7B model of Show-o2 using DPG-Bench [ 23 ] . The subscript values show the speedup relative to Show-o2 baseline. Latency is measured in seconds (s).
Strategy
Latency (s)
Score
Show-o2 [ 72 ]
5.23
85.02
FlashU (Ours)
2.80 +1.87 ×
84.12
w/o Dynamic Layer Pruning
2.91 +1.80 ×
85.33
w/o Hybrid Network
2.48 +2.11 ×
83.57
w/o Adaptive Guidance
3.92 +1.33 ×
83.42
FFN Pruning Only
3.84 +1.36 ×
84.27
Table 5: Ablation study on the 1.5B model of Show-o2 using DPG-Bench [ 23 ] . The subscript values show the speedup relative to Show-o2 baseline. Latency is measured in seconds (s).
Method
# Params.
Global
Entity
Attribute
Relation
Overall ↑
Latency (s) ↓
Hunyuan-DiT [ 34 ]
1.5B
84.59
80.59
88.01
74.36
78.87
20.1
Playground v2.5 [ 32 ]
-
83.06
82.59
81.20
84.08
75.47
7.67
TokenFlow-t2i [ 48 ]
7B
72.64
74.72
77.67
83.80
67.44
0.97
JanusFlow [ 46 ]
1.5B
81.16
85.16
85.32
91.05
77.57
1.09
VILA-U [ 70 ]
7B
89.92
82.73
80.80
87.83
74.26
9.78
Liquid [ 69 ]
8B
79.64
87.61
85.24
91.01
80.26
37.05
Table 6: Latency Comparison on DPG-Bench [ 23 ] . Results in gray correspond to vanilla Show-o2. Subscripts indicate the speedup of FlashU. Latency is reported in seconds (s), averaged per sample.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Latency vs. Performance Trade-off. Comparison of FlashU against native unified models (Show-o2, VILA-U, Emu3-DPO, Liquid) and tailored models (JanusFlow, TokenFlow-t2i). The x-axis represents inference latency (lower is better), and the y-axis represents overall performance on DPG-Bench(higher is better). Circle sizes denote model parameter counts (1.5B, 7B, 8B).
Hyperparameter
Symbol
Value
Total Inference Steps
T
0.7T0
Hybrid Threshold
τ
0.3
FFN Pruning Ratio
rp
0.20
Layer Skipping Ratio
rLS
0.20
Recalculation Interval
TLS
10
Diffusion Head Cache Interval
Tcache
5
Appendix
Table 7: Hyperparameters for Generation Tasks.
Hyperparameter
Symbol
Value
Max Generated Tokens
Nmax
128
Hybrid Threshold
τ
0.3
FFN Pruning Ratio
rp
0.10
Layer Skipping Ratio
rLS
0.10
Calibration Samples
Ncalib
20
Recalculation Interval
TLS
10
Appendix
Table 8: Hyperparameters for Multimodal Understanding Tasks.
Figure 8: Sensitivity Analysis of FFN Pruning Ratio. The trade-off between generation quality (DPG-Bench score) and inference speedup is shown as the FFN pruning ratio ( rp ) varies from 0.1 to 0.5.
Figure 9: Sensitivity Analysis of Layer Skipping Ratio. The relationship between generation quality and speedup across different layer skipping ratios ( rLS ).
Figure 10: Comparative Analysis of Hybrid Threshold Strategies. The efficiency-quality trade-off curves for Adaptive Guidance versus Baseline strategies across different hybrid threshold ( τ ) values.
Figure 11: Component-wise Latency Profiling Analysis. Inference time breakdown by component for Show o2 and FlashU on the 1.5B model.
Figure 12: Qualitative results of text-to-image generation. We showcase samples generated by FlashU across diverse prompts. The model successfully handles complex descriptions, rendering high-quality textures and correct object placements, demonstrating that our acceleration framework preserves the generative performance of the backbone model.
Unified multimodal models jointly support understanding and generation, but incur substantial redundant computation across tokens, layers, and generation timesteps. Through token-importance probing, we identify an asymmetric core-expansion structure: understanding exhibits a stable importance component, while generation largely shares this component but requires progress-dependent corrections. We therefore propose CE-Router, which uses a task-shared core scorer and progress-conditioned generation expansions, optimized through generation decomposition and cross-task core alignment. At inference, CE-Router compacts token computation and supplies a learned routing signal to Unified Computation Scheduling, which coordinates layer skipping, FFN pruning, diffusion-head cache reuse, and denoising-step early exit. Experiments on two representative UMM architectures demonstrate consistent quality--efficiency improvements across both tasks, retaining 98.03% of dense understanding performance with a 1.93× end-to-end inference speedup.
We present Lance, a lightweight native unified model supporting multimodal understanding, generation, and editing for both images and videos. Rather than relying on model capacity scaling or text-image-dominant designs, Lance explores a practical paradigm for unified multimodal modeling via collaborative multi-task training. It is grounded in two core principles: unified context modeling and decoupled capability pathways. Specifically, Lance is trained from scratch and employs a dual-stream mixture-of-experts architecture on shared interleaved multimodal sequences, enabling joint context learning while decoupling the pathways for understanding and generation. We further introduce modality-aware rotary positional encoding to mitigate interference among heterogeneous visual tokens and boost cross-task alignment. During training, Lance adopts a staged multi-task training paradigm with capability-oriented objectives and adaptive data scheduling to strengthen both semantic comprehension and visual generation performance. Experimental results demonstrate that Lance substantially outperforms existing open-source unified models in image and video generation, while retaining strong multimodal understanding capabilities. The homepage is available at https://lance-project.github.io.
Unified models aim to support both understanding and generation by encoding images into discrete tokens and processing them alongside text within a single autoregressive framework. This unified design offers architectural simplicity and cross-modal synergy, which facilitates shared parameterization, consistent training objectives, and seamless transfer between modalities. However, the large number of visual tokens required by such models introduces substantial computation and memory overhead, and this inefficiency directly hinders deployment in resource constrained scenarios such as embodied AI systems. In this work, we propose a unified token compression algorithm UniCompress that significantly reduces visual token count while preserving performance on both image understanding and generation tasks. Our method introduces a plug-in compression and decompression mechanism guided with learnable global meta tokens. The framework is lightweight and modular, enabling efficient integration into existing models without full retraining. Experimental results show that our approach reduces image tokens by up to 4 times, achieves substantial gains in inference latency and training cost, and incurs only minimal performance degradation, which demonstrates the promise of token-efficient unified modeling for real world multimodal applications.