Unified models (UMs) hold promise for their ability to understand and generate content across heterogeneous modalities. Compared to merely generating visual content, the use of UMs for interleaved cross-modal reasoning is more promising and valuable, e.g., for solving understanding problems that require dense visual thinking, improving visual generation through self-reflection, or modeling visual dynamics of the physical world guided by stepwise action interventions. However, existing UMs necessitate pixel decoding as a bridge due to their disjoint visual representations for understanding and generation, which is both ineffective and inefficient. In this paper, we introduce LatentUM, a novel unified model that represents all modalities within a shared semantic latent space, eliminating the need for pixel-space mediation between visual understanding and generation. This design naturally enables flexible interleaved cross-modal reasoning and generation. Beyond improved computational efficiency, the shared representation substantially alleviates codec bias and strengthens cross-modal alignment, allowing LatentUM to achieve state-of-the-art performance on the Visual Spatial Planning benchmark, push the limits of visual generation through self-reflection, and support world modeling by predicting future visual states within the shared semantic latent space.
Figures & tables
Figure 1: Latent-space unified models enable interleaved cross-modal reasoning through shared semantic visual representations. Left: LatentUM improves text-to-image generation via self-reflection over its own generated semantic visual tokens. Middle: LatentUM interleaves textual reasoning with latent visual state updates for visual spatial planning. Right: LatentUM supports world modeling by predicting future visual states as semantic tokens conditioned on actions.
Figure 2: Overview of LatentUM. LatentUM unifies visual understanding and generation within a shared semantic latent space, enabling cross-modal reasoning without pixel-space mediation. (a) The visual tokenizer Qϕ is trained via model behavior aligned quantization (MBAQ): it minimizes the KL divergence between the VLM’s output distributions on original features V versus quantized features Qϕ(V) , preserving understanding-oriented semantics rather than pixel details. (b) Mixture-of-Modal Experts (MoME): the transformer maintains two parallel branches, where ψ handles understanding and θ handles generation. The two branches share one self-attention mechanism. Generated visual codes are de-quantized and recached in the same context, allowing the model to reason over its own outputs. (c) Decoupled pixel decoder: a diffusion decoder ϵη optionally maps quantized semantic features to pixels for visualization. It is trained independently, so the core model never optimizes for pixel fidelity.
Figure 3
Model
MME-P ↑
POPE ↑
SEED ↑
MMBench ↑
MMMU val ↑
Tokenflow XL ( Qu et al., 2025 )
1551
87.8
72.6
76.8
43.2
Bagel ( Deng et al., 2025 )
1687
-
-
85.0
55.3
JanusFlow ( Ma et al., 2025b )
1333
88.0
70.5
-
-
Harmo ( Wu et al., 2025d )
1155
87.6
67.1
65.5
-
Chameleon ( Team, 2024 )
1057
77.8
-
-
26.7
Orthus ( Kou et al., 2024 )
1266
79.6
-
-
28.2
Table 1: Evaluation on multimodal understanding benchmarks. LatentUM Base achieves competitive performance while preserving the pretrained VLM’s visual understanding capability. † denotes the quantized setting where LatentUM Base processes benchmark images through quantized visual representations.
Figure 5: Text-to-image gallery of LatentUM. The last column highlights LatentUM’s text rendering capability, which emerges as visual and language tokens share a unified semantic space, enabling legible in-image text.
Model
Single Obj.
Two Obj.
Counting
Colors
Position
Color Attr.
Overall ↑
Specialized Text-to-Image Models
FLUX.1 [Dev] ( Labs, 2024 )
0.98
0.81
0.74
0.79
0.22
0.45
0.66
SD3.5 Large ( Esser et al., 2024 )
0.98
0.89
0.73
0.83
0.34
0.47
0.71
Qwen-Image ( Wu et al., 2025a )
0.99
0.92
0.89
0.88
0.76
0.77
0.87
Unified Models
TokenFlow-XL ( Qu et al., 2025 )
0.95
0.60
0.41
0.81
0.16
0.24
0.55
Table 2: Evaluation on GenEval ( Ghosh et al., 2023 ) . LatentUM achieves the best performance among unified models on compositional instruction-following for visual generation. LatentUM Vis-Gen is post-trained from LatentUM Base via GRPO with self-derived reward; pixel-reward denotes further replacing the reward signal with the external GenEval pixel-reward model.
TIFA GM
TIFA AM
Bagel + CoT ( Deng et al., 2025 )
23.1
70.9
Janus Pro ( Chen et al., 2025b )
14.5
58.3
BLIP3o 8B ( Chen et al., 2025a )
13.3
59.4
LatentUM Base
21.4
68.1
LatentUM Vis-Gen
31.3
72.9
Table 3: Evaluation on GenEval2 ( Kamath et al., 2025 ) . LatentUM Vis-Gen outperforms other unified models by a large margin.
Figure 6: Two interleaved reasoning paradigms for visual spatial planning. (a) Coarse-grained: LatentUM Vis-Plan first analyzes the maze layout, generates a complete visual plan, then derives the solution. (b) Fine-grained: LatentUM Vis-Plan alternates between textual reasonings and visual state updates, where each generated visual state serves as context for the next action.
Model
Lv.3
Lv.4
Lv.5
Lv.6
Avg.
Anole ( Chern et al., 2024 )
0.02
0.01
0.00
0.00
0.01
MVoT ( Li et al., 2025a )
0.21
0.11
0.08
0.03
0.11
Mirage ( Yang et al., 2025 )
0.93
0.83
0.76
0.51
0.76
Chameleon ( Team, 2024 )
-
-
-
-
0.01
InternVL3.5 8B ( Wang et al., 2025a )
-
-
-
-
0.08
InternVL3.5 38B ( Wang et al., 2025a )
-
-
-
-
0.20
Table 4: Evaluation on the VSP benchmark ( Wu et al., 2024a ) . LatentUM Vis-Plan outperforms all prior visual reasoning models under both planning paradigms, with the fine-grained variant achieving near-perfect accuracy.
Figure 7: World Modeling for Visual Navigation. (a) Open-loop future frame prediction. Given initial context frames (red box) and a predefined action trajectory, LatentUM WM rolls out temporally consistent future observations that preserve the spatial layout of the initial scene. (b) Zero-shot language-conditioned navigation. Without language supervision during fine-tuning, LatentUM WM translates free-form textual instructions (e.g., “get closer to the left house”) into accurate future visual states, enabled by the inherited language-vision alignment. (c) Quantitative comparison on the NWM benchmark ( Bar et al., 2025 ) . LatentUM WM surpasses the unified baseline Transfusion-RAE on both ATE and RPE, and remains competitive with the specialized NWM model.
Figure 8: Ablation study. (a) Text-to-image generation by the VQVAE baseline using the same prompts as Figure 5 (leftmost and rightmost of the first two rows). (b) Performance comparison on GenEval and GenEval2. Semantically aligned features induce stronger instruction-following capability during visual generation. (c) Comparison of semantic feature quantizers trained with different objectives on visual understanding benchmarks. MBAQ consistently outperforms feature-reconstruction baselines.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: Generation results under different CFG scales ( 1.0 to 5.0 ) for identical prompts. LatentUM produces coherent images even at CFG =1.0 .
Figure 10: Reward curve during GRPO training for self-reflection visual generation.
Figure 11: Qualitative comparison before and after self-reward GRPO. Top: LatentUM Base before GRPO. Bottom: the model after self-reward GRPO. The post-trained model follows counting prompts more accurately.
Figure 12: Examples where the generated images are qualitatively correct, but the detection-based metrics mark them as failures.
Figure 13: Top: independent keyframe decoding with the original decoupled decoder. Bottom: reference-conditioned decoding with the first-frame VAE latent. The bottom row shows much better cross-frame consistency.
Unified multimodal models (UMMs) achieve strong performance in both understanding and generation by learning a shared latent space, yet they often exhibit functional inconsistency between these two capabilities. We observe that this issue does not stem from a lack of shared representations, but from the absence of explicit alignment between the transformations that map into and out of the latent space. As a result, generation and re-encoding can follow inconsistent trajectories, leading to semantic drift under modality transitions. In this work, we propose LatentUMM, a framework that constructs an enhanced shared latent space to explicitly align these transformations and improve cross-modal consistency. LatentUMM consists of two stages. First, dual latent alignment enforces consistency at both the modality and capacity levels: cross-modal alignment uses a stronger embedding model to impose structured cross-modal semantics, while dual capacity alignment enforces bidirectional consistency under generation and re-encoding. Second, latent dynamics stabilization improves robustness via stochastic latent rollouts and preference optimization, favoring trajectories that better preserve semantic consistency. Experiments show that LatentUMM consistently improves multimodal consistency across diverse architectures. Code is available at: https://github.com/AIFrontierLab/TorchUMM/tree/main/src/umm/post_training/LatentUMM.
While unified multimodal models (UMMs) jointly perform visual understanding and generation within a single model, functional unification does not guarantee learning synergy: the two objectives may reinforce each other, compete for capacity, or merely coexist. We investigate their relationship at the representation, task, and system levels in a controlled, structurally native setting without pretrained vision priors. At the representation level, we find that each objective provides useful signal to the other: generation enriches the visual features learned for understanding, while understanding strengthens vision--language alignment for generation. However, when both objectives are forced through the same computation path, one tends to dominate. A task-decoupled architecture that specializes conflicting visual computation while preserving semantic interaction avoids this asymmetric degradation. At the task level, through three case studies, we find positive bidirectional transfer when understanding and generation tasks rely on shared knowledge. At the system level, we show that an end-to-end UMM outperforms a matched planner--executor pipeline on complex tasks that explicitly require both image understanding and generation. Together, these results show that the value of UMMs extends beyond a unified interface: appropriate specialization, shared task knowledge, and end-to-end optimization can turn coexistence into synergy.
Penghao Wu, Haiwen Diao, Weichen Fan +3
1S-Lab, Nanyang Technological University · 2SenseTime Research
Unified multimodal models (UMMs) that integrate understanding, reasoning, generation, and editing face inherent trade-offs between maintaining strong semantic comprehension and acquiring powerful generation capabilities. In this report, we present InternVL-U, a lightweight 4B-parameter UMM that democratizes these capabilities within a unified framework. Guided by the principles of unified contextual modeling and modality-specific modular design with decoupled visual representations, InternVL-U integrates a state-of-the-art Multimodal Large Language Model (MLLM) with a specialized MMDiT-based visual generation head. To further bridge the gap between aesthetic generation and high-level intelligence, we construct a comprehensive data synthesis pipeline targeting high-semantic-density tasks, such as text rendering and scientific reasoning, under a reasoning-centric paradigm that leverages Chain-of-Thought (CoT) to better align abstract user intent with fine-grained visual generation details. Extensive experiments demonstrate that InternVL-U achieves a superior performance - efficiency balance. Despite using only 4B parameters, it consistently outperforms unified baseline models with over 3x larger scales such as BAGEL (14B) on various generation and editing tasks, while retaining strong multimodal understanding and reasoning capabilities.
Changyao Tian, Danni Yang, Guanzhou Chen +26
Shanghai AI Laboratory · CUHK MMLab · Fudan University +6