Paper deep dive
UniSpace: Unified Visual Representation and Scalable Multimodal Modeling
Jinbo Yan, Limeng Qiao, Jie Qin, Junyan He, Feize Wu, Guanglu Wan
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 8/12/2026, 2:20:32 AM
Summary
The paper introduces UniSpace, an 8B parameter Mixture-of-Transformer-Experts model that unifies visual understanding, generation, and editing within a single visual representation space. It addresses the limitation of semantic vision encoders (ViTs) discarding fine-grained details by introducing 'Patch Reparameterization,' which adds a reconstruction-aware patch embedding to a frozen semantic ViT. This allows the model to preserve semantic pathways while enabling high-fidelity image reconstruction and generation without a separate VAE pathway.
Entities (10)
Relation Signals (10)
Patch Reparameterization â adds â reconstruction-aware patch embedding
confidence 95% ¡ introduce Patch Reparameterization, which preserves the original semantic pathway while adding a reconstruction-aware patch embedding
UniSpace â performs â Generation
confidence 95% ¡ performs understanding, generation, and editing in the same visual space
UniSpace â performs â editing
confidence 95% ¡ performs understanding, generation, and editing in the same visual space
UniSpace â performs â Understanding
confidence 95% ¡ performs understanding, generation, and editing in the same visual space
UniSpace â uses â Patch Reparameterization
confidence 95% ¡ We scale PR-Qwen-ViT into UniSpace... showing that a reparameterized pretrained ViT can serve as a unified visual interface
PR-Qwen-ViT â isbasefor â UniSpace
confidence 90% ¡ We scale PR-Qwen-ViT into UniSpace
SigLIP2 â isbasefor â PR-SigLIP2
confidence 90% ¡ Applying this design to several pretrained encoders gives PR-SigLIP2
DinoV2 â isbasefor â PR-DINOv2
confidence 90% ¡ Applying this design to several pretrained encoders gives PR-SigLIP2, PR-DINOv2
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Semantic vision encoders have become a central visual interface for multimodal understanding and semantic conditioning in image generation. However, their final tokens discard fine-grained visual details, leading to poor pixel reconstruction and limiting their use in reconstruction-sensitive tasks such as image generation and editing. In this work, we ask whether understanding, generation, and editing can be modeled in a single visual representation space built from a pretrained semantic ViT. We show that the frozen Transformer blocks of a semantic ViT are not intrinsically unable to preserve visual details. Instead, the original patch parameterization drives the representation toward semantic abstraction, making fine-grained information difficult to recover from the final tokens. Based on this observation, we introduce \emph{Patch Reparameterization}, which preserves the original semantic pathway while adding a reconstruction-aware patch embedding that provides fine-grained visual information to the same frozen ViT blocks. The resulting unified representation preserves multimodal understanding while enabling high-fidelity image reconstruction and a favorable reconstruction--generation trade-off. We further scale this representation into \emph{UniSpace}, an 8B Mixture-of-Transformer-Experts model that performs understanding, generation, and editing in the same visual space without a separate VAE pathway. System-level evaluations demonstrate practical text-to-image generation and instruction-based image editing, showing that a reparameterized pretrained ViT can serve as a unified visual interface for scalable multimodal modeling.
Tags
Links
- Source: https://arxiv.org/abs/2608.08676v1
- Canonical: https://arxiv.org/abs/2608.08676v1
Trouble viewing inline? Open PDF directly â
Full Text
93,676 characters extracted from source content.
Expand or collapse full text
UniSpace UniSpace: Unified Visual Representation and Scalable Multimodal Modeling Jinbo YanLimeng QiaoJie QinJunyan He Feize WuGuanglu Wan Meituan ABSTRACT Semantic vision encoders have become a central visual interface for multimodal understanding and semantic conditioning in image generation. However, their final tokens discard fine-grained vi- sual details, leading to poor pixel reconstruction and limiting their use in reconstruction-sensitive tasks such as image generation and editing. In this work, we ask whether understanding, gener- ation, and editing can be modeled in a single visual representation space built from a pretrained semantic ViT. We show that the frozen Transformer blocks of a semantic ViT are not intrinsically unable to preserve visual details. Instead, the original patch parameterization drives the represen- tation toward semantic abstraction, making fine-grained information difficult to recover from the final tokens. Based on this observation, we introduce Patch Reparameterization, which preserves the original semantic pathway while adding a reconstruction-aware patch embedding that provides fine-grained visual information to the same frozen ViT blocks. The resulting unified representation preserves multimodal understanding while enabling high-fidelity image reconstruction and a favor- able reconstructionâgeneration trade-off. We further scale this representation into UniSpace, an 8B Mixture-of-Transformer-Experts model that performs understanding, generation, and editing in the same visual space without a separate VAE pathway. System-level evaluations demonstrate practi- cal text-to-image generation and instruction-based image editing, showing that a reparameterized pretrained ViT can serve as a unified visual interface for scalable multimodal modeling. Project: https://yjb6.github.io/UniSpace/ Hugging Face: https://huggingface.co/yjb6/UniSpace GitHub: https://github.com/yjb6/UniSpace 0.53 0.53 0.53 0.53 0.53 3.90 3.20 4.28 87.8 85.1 86.5 54.9 36.1 56.1 53.5 37.0 53.3 ImgEditDPGOneIG-Bench(EN)OneIG-Bench(CN) SenseNova-U1UniSpace(Ours) Bagel 1 arXiv:2608.08676v1 [cs.CV] 9 Aug 2026 UniSpace CONTENTS 1 Introduction3 2 Related Work4 2.1Semantic Representations for Generation and Reconstruction . . . . . . . . . . . . . . . . . . . . . .4 2.2Visual Representation Spaces in Unified Multimodal Models . . . . . . . . . . . . . . . . . . . . . .4 3 Unified Encoder via Patch Reparameterization4 3.1Patch Reparameterization Can Carry Visual Details . . . . . . . . . . . . . . . . . . . . . . . . . . .5 3.2Constructing a Unified Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 3.3Training for Reconstruction and Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6 3.4Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7 4 UniSpace: Scaling Unified Modeling in One Visual Space8 4.1Unified MoT Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8 4.2Unified Training Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9 4.3Training Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10 5 Experiments11 5.1Unified Tokenizer Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11 5.1.1High-Fidelity Image Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11 5.1.2Multimodal Understanding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .13 5.1.3ImageNet Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .14 5.1.4Auxiliary Diagnostic Study on Entangled Unified Representations . . . . . . . . . . . . . . .15 5.2Scaling Unified Multimodal Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .16 5.2.1Image Editing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .16 5.2.2Text-to-Image Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .17 5.2.3Unified Capability Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .20 6 Limitation20 7 Conclusion23 A Ablation on Balanced Flow Matching28 B Ablation on Reconstruction Compression28 C Entangled Representation and Diagnostic Protocol28 2 UniSpace (a) A one-layer swap recovers +3.70 dB 0369 Last 20 25 30 35 40 0 20 40 60 80 ViT layer PSNR (dB) Accuracy (%) +3.70 dB Pretrained P s Random P rand Acc. of P s (b) Same frozen blocks, different information Pretrained P s Semantic-selective transformations Semantic-dominant deep features identical frozen Transformer weights pretrained route suppresses low-level variation Random P rand Semantic response disrupted Detail-richer deep features The patch embedding is the bottleneck: frozen blocks can still carry pixel detail. Figure 1: The patch embeddingânot the frozen Transformer blocksâis the reconstruction bottleneck. (a) As semantic accuracy rises with depth, pixel recoverability through the pretrained SigLIP2 pathway collapses. Replac- ing only its patch embedding P s with a random projection P rand raises last-layer PSNR by 3.70 dB; all Transformer weights remain frozen and the reconstruction probes are identical. (b) This controlled intervention shows that input parameterization determines which information route the same backbone activates: the pretrained route favors seman- tic abstraction, whereas disrupting it exposes a detail-richer residual pathway. 1INTRODUCTION Semantic vision encoders have become one of the most important visual interfaces in modern multimodal systems. In vision-language models, encoders such as SigLIP (Tschannen et al., 2025) and CLIP (Radford et al., 2021) pro- vide semantically aligned visual tokens for image understanding. Beyond understanding, they are also widely used in image generation and editing systems as semantic conditioning signals. However, these tokens are primarily opti- mized for semantic abstraction and alignment, rather than for preserving fine-grained visual details. Their final-layer representations often exhibit poor pixel-level recoverability, making them insufficient as the sole visual interface for detail-sensitive tasks such as image generation and reference-preserving editing. This limitation leads current generation and editing systems to rely on fragmented visual representations. Text-to- image models typically generate images in an autoencoding latent space, such as VAE latents, while semantic encoders are introduced separately when high-level image semantics are needed. The separation becomes more problematic for image editing, where the reference image must provide both semantic guidance, such as object identity and layout, and fine-grained details, such as texture, color, and local structure. As a result, existing systems often combine semantic encoder features with reconstruction-oriented latents or features. Recent unified multimodal models (Deng et al., 2025) inherit the same issue: although understanding, generation, and editing may be placed within a single backbone, the visual information is still represented through separate semantic and reconstruction-oriented spaces. They are unified at the architecture level, but not at the level of visual representation. This work asks whether visual understanding, generation, and editing can be modeled within a single visual rep- resentation space built upon a pretrained semantic ViT. This raises two central challenges. First, we need a visual representation that unifies semantic understanding and fine-grained detail reconstruction while retaining the pretrained semantic backbone. Second, we need to scale this representation to an LLM-based unified multimodal model that supports understanding, generation, and editing in the same visual space. To address the first challenge, we revisit why semantic ViTs reconstruct poorly. As shown in Fig. 1, the bottleneck lies not in the frozen Transformer blocks, but in the original semantic patch embedding, which suppresses fine-grained visual details. This motivates Patch Reparameterization: we retain the original patch embedding for semantic under- standing and add a reconstruction-aware embedding to supply visual details to the same frozen backbone. Merging the two token streams yields a unified representation for both understanding and reconstruction, which we train with a bal- anced flow-matching objective. Applying this design to several pretrained encoders gives PR-SigLIP2, PR-DINOv2, and PR-Qwen-ViT, which preserve semantic understanding while enabling high-fidelity reconstruction, generation, and reference-preserving image editing. To address the second challenge, we scale PR-Qwen-ViT into UniSpace, an 8B Qwen-based Mixture-of-Transformer- Experts multimodal model. UniSpace uses the patch-reparameterized ViT encoderâdecoder as its sole visual interface. 3 UniSpace Unlike prior unified models that combine a semantic vision encoder with a separate VAE latent space, UniSpace represents reference images, target images, and generated outputs in the same unified visual representation space. This design directly tests whether a reparameterized pretrained semantic ViT can serve as a shared visual interface for understanding, generation, and editing at scale. Our experiments show that UniSpace enables strong text-to-image generation and instruction-based image editing while retaining general visual understanding capabilities under unified generative training. Our contributions are threefold: ⢠We introduce Patch Reparameterization, showing that frozen pretrained semantic ViTs can preserve fine- grained visual information and can be adapted to provide both semantic and reconstruction-aware visual tokens within a single backbone. ⢠We build a family of unified visual encoders, including PR-SigLIP2, PR-DINOv2, and PR-Qwen-ViT, which provide a balanced visual representation for understanding, reconstruction, and generation. ⢠We develop UniSpace, a Qwen-based unified multimodal model powered by PR-Qwen-ViT, where a sin- gle visual encoderâdecoder interface supports multimodal understanding, text-to-image generation, and instruction-based image editing. 2RELATED WORK 2.1SEMANTIC REPRESENTATIONS FOR GENERATION AND RECONSTRUCTION Recent works have shown that semantic representations are highly beneficial for image generation. REPA (Yu et al., 2024) aligns intermediate diffusion-transformer features with pretrained semantic representations to accelerate con- vergence and improve generation quality. VA-VAE (Yao et al., 2025b) introduces semantic alignment into autoencoder training, making the latent space easier for generative models to learn. RAE (Zheng et al., 2025) further demonstrates that frozen semantic encoder features can directly serve as generative latents when paired with a learned reconstruction decoder. Nevertheless, semantic representations are not automatically unified visual tokenizers. Such tokenizers must jointly support semantic understanding, high-fidelity reconstruction, and generative modeling. UniFlow (Yue et al., 2025) constructs understanding-aware reconstruction representations through distillation, and RAEv2 (Singh et al., 2026) improves RAE-style reconstruction by incorporating multi-layer DINO (Caron et al., 2021)features. These works extend semantic representations toward reconstruction, but the balance among semantic capability, reconstruction fidelity, and generative modeling remains challenging. In particular, strong understanding and reconstruction do not necessarily imply a representation that is easy for a generative prior to model. Patch Reparameterization targets this gap by improving reconstruction fidelity within a frozen semantic ViT while keeping the resulting representation suitable for generative modeling. 2.2VISUAL REPRESENTATION SPACES IN UNIFIED MULTIMODAL MODELS Recent works have explored unified multimodal models that support understanding, generation, and editing within a single framework (Cui et al., 2025; Wang et al., 2024; SenseNova-U1 Team, 2026; Deng et al., 2025). However, they differ substantially in how visual information is represented. Some methods model images with discrete visual tokenizers (Cui et al., 2025; Wang et al., 2024), while others operate in continuous latent spaces. A common design is to use separate pretrained visual spaces for different tasks. For example, BAGEL (Deng et al., 2025) adopts a Mixture- of-Transformer-Experts architecture in which semantic vision-encoder tokens support understanding and VAE latents support generation, with shared self-attention enabling interaction between the two pathways. Although the model is unified architecturally, its visual representation remains split across semantic and reconstruction-oriented spaces. Our UniSpace instead preserves the semantic prior of a pretrained ViT and uses the patch-reparameterized encoderâ decoder as the only visual tokenizer. This allows understanding, generation, and editing to operate in the same unified visual representation space while benefiting from the efficiency and transferability of pretrained semantic encoders, without introducing a separate VAE latent space. 3UNIFIED ENCODER VIA PATCH REPARAMETERIZATION Pretrained semantic ViTs, such as SigLIP and CLIP, provide strong visual perception and image-text alignment, but their final tokens are not designed for faithful pixel reconstruction. Existing attempts to obtain understanding-aware 4 UniSpace reconstruction representations often train a new visual encoder (Yao et al., 2025a) or rely on semantic distillation (Yue et al., 2025), which complicates the system and can alter the original semantic representation. We ask a simpler question: is it necessary to relearn the Transformer blocks, or can a frozen semantic ViT be reparameterized to expose the visual details that its original semantic tokens suppress? We answer this question with Patch Reparameterization. The pretrained patch embedding is kept unchanged to pre- serve the original semantic pathway, while an additional reconstruction-aware patch embedding provides an alternative parameterization that preserves reconstruction-relevant visual information within the same frozen Transformer blocks. Through a simple Token Fusion strategy, where reconstruction tokens are compressed and concatenated with semantic tokens, the resulting representation unifies semantic and reconstruction information, providing a shared visual interface for understanding and generative modeling. 3.1PATCH REPARAMETERIZATION CAN CARRY VISUAL DETAILS Prior analyzes (Singh et al., 2026) have suggested that shallow features in semantic ViTs preserve more fine-grained visual details, whereas deeper features become more semantically abstract. Consistent with this observation, Fig. 1 shows that as visual tokens propagate through deeper Transformer blocks, their semantic classification accuracy in- creases from 0.93% to 81.29%, while their pixel-level recoverability degrades rapidly. However, this degradation does not necessarily imply that the frozen Transformer blocks are intrinsically unable to carry visual details. Unlike autoen- coders with an explicit low-dimensional bottleneck, ViT backbones maintain high-dimensional token representations throughout the network. This design provides a plausible pathway for propagating input-dependent details across layers. To test whether the poor recoverability is imposed by the frozen Transformer blocks themselves, we conduct a diagnos- tic experiment on SigLIP2. We replace its pretrained patch embedding with a randomly initialized linear projection, keep all subsequent Transformer blocks frozen, and train identical reconstruction probes on features extracted at dif- ferent depths. If the frozen semantic Transformer blocks were incapable of carrying pixel-level information, changing only the input projection would not improve recoverability at deep layers. Counterintuitively, the random patch embedding improves the last-layer PSNR from 20.96 to 24.66, despite using exactly the same frozen Transformer blocks. As shown in Fig. 1, at the patch-embedding output, the pretrained and random projections are similarly recoverable, achieving PSNR values of 39.29 and 39.68, respectively. Their behavior diverges only as the tokens pass through the same frozen semantic Transformer blocks. Under the pretrained patch embedding, the input activates visual patterns that the subsequent blocks were optimized to semantically abstract, progressively suppressing variations irrelevant to the semantic objective. The random projection disrupts this preferred semantic processing trajectory, allowing more low-level input variation to remain recoverable through the residual pathway. This result suggests that pixel-level information is not lost because the hidden space cannot carry it. Rather, it is selectively suppressed under the semantic encoderâs original input parameterization. However, randomizing the patch embedding is not itself a useful solution, since it destroys the pretrained semantic representation. It instead moti- vates a minimal adaptation: keep the original semantic patch embedding for understanding, and learn an additional reconstruction-aware patch embedding that injects visual details into the same frozen Transformer blocks, as shown in Fig.2. 3.2CONSTRUCTING A UNIFIED REPRESENTATION Given an input image I , we construct a unified representation that preserves the semantic ability of the pretrained encoder while adding reconstruction details in a channel-factorized form. Let F Ď denote the frozen Transformer blocks of a pretrained semantic ViT, and let P s denote its original pretrained patch embedding. The original semantic tokens are T s = F Ď (P s (I)),(1) where both P s and F Ď are kept frozen. This path preserves the pretrained semantic representation and provides the basis for visual understanding. To capture reconstruction-relevant details, we introduce a reconstruction-aware patch embedding P r , initialized from P s and optimized for reconstruction. It is followed by the same frozen Transformer blocks: T r = F Ď (P r (I)).(2) 5 UniSpace Semantic Patch Embedding Unified Representation Understanding Semantic ViT Blocks Understanding Q: Whatâs this? A:A parrot Reconstruction Generation : Trainable: Frozen Transformer Blocks N â Reconstruction Recognition Q: Whatâs this? A: A parrot â Reconstrution â A parrot â Patch Embedding N Transformer Blocks Transformer Blocks Patch Embedding Transformer Blocks TokenFusion Patch EmbeddingLayer Add a New Patch Embedding Original Semantic Encoder Patch-reparameterized Unified Encoder Patch EmbeddingLayer Result: Figure 2: Overview of Patch Reparameterization. The original semantic patch embedding and frozen ViT blocks preserve the pretrained semantic pathway, while a reconstruction-aware patch embedding injects visual details into the same frozen backbone. The semantic and reconstruction-aware tokens are combined by a Token Fusion layer to form the unified representation, which is shared by understanding, reconstruction, and generation. Thus, T r is not produced by an independent reconstruction encoder; it is encoded within the same semantic ViT parameter space through a different input parameterization. Explicit Token Fusion. A naive way to unify T s and T r is to fuse them into a single entangled feature space using an MLP. However, as discussed in Sec. 5.1.4, this implicit fusion makes it difficult to control the relative contributions of semantic and reconstruction information during generative modeling. We therefore preserve the two components as explicitly separated feature groups along the channel dimension. Before concatenation, we first project the recon- struction tokens into a compact space: e T r = W r T r ,(3) where W r is a learnable linear projection along the channel dimension, mapping the reconstruction tokens from d to a lower dimension d r . This projection serves a distinct role from concatenation: it reduces the dimensionality of the reconstruction component, lowering the burden for the generative model while retaining decoder-critical visual details. We then concatenate the original semantic tokens and the compressed reconstruction tokens along the channel dimension: T u = Concat(T s , e T r ).(4) The complete unified representation T u serves as the shared visual representation for all downstream tasks, including understanding, reconstruction, and generation; neither T s nor e T r is used independently. Unlike learned feature merg- ing, this explicit fusion leaves the pretrained semantic representation unchanged and maintains an explicit boundary between semantic and reconstruction information. Compression reduces the dimensionality of reconstruction details, while concatenation makes the two information sources separately addressable during generative training. 3.3TRAINING FOR RECONSTRUCTION AND GENERATION Training the Reconstruction-Aware Path. After constructing T u , we adopt a ViT decoder D Ď , following RAE, to reconstruct the input image: Ë I = D Ď (T u ).(5) During training, the reconstruction-aware components and the decoder are optimized with a reconstruction objective: L rec =D( Ë I,I),(6) 6 UniSpace where D is instantiated with pixel-level and perceptual reconstruction losses. Only the reconstruction-aware patch embedding P r , the linear projection W r , and the decoder D Ď are updated. The original patch embedding P s and the pretrained ViT blocks F Ď remain frozen. Since reconstruction training never modifies the original semantic path, the pretrained semantic tokens are structurally preserved while the additional path learns to supply decoder-critical details. Balanced Generative Modeling. The explicit channel-wise decomposition of T u also gives direct control over gen- erative training. A plain MSE over the concatenated representation weights the two components according to their di- mensionality and scale, without specifying how much of the objective should focus on decoder-critical reconstruction information. We therefore use a balanced flow-matching objective that normalizes each component by its dimension and then assigns an explicit objective weight. We model the distribution of the unified representation using conditional flow matching. Let Z 1 = T u denote a unified representation, Z 0 âź N(0,I) a Gaussian noise sample of the same shape, and t âź U[0, 1] a timestep. We construct the linear probability path Z t = (1â t)Z 0 + tZ 1 ,(7) whose target velocity is V t = dZ t dt = Z 1 â Z 0 .(8) Given Z t , t, and condition c, the generative model v θ predicts the velocity b V t = v θ (Z t ,t,c). Let T s â R NĂa and e T r â R NĂb denote the semantic and reconstruction components of T u , respectively. Their explicit channel-wise separation allows us to partition the target and predicted velocities as V t = Concat(V s t ,V r t ), b V t = Concat( b V s t , b V r t ).(9) We normalize the prediction error of each component by its number of dimensions and optimize the reconstructionâ semantic balanced flow-matching objective L BFM = E Z 1 ,Z 0 ,t,c " (1â Îť r ) ⼠b V s t â V s t ⼠2 F Na + Îť r ⼠b V r t â V r t ⼠2 F Nb # ,(10) where Îť r specifies the fraction of the total objective weight assigned to the reconstruction component. We set Îť r = 0.75, allocating 75% of the objective weight to the decoder-critical reconstruction component while retaining the semantic component that facilitates generative learning; this choice is supported by the ablation in Table 12. At inference time, we sample Z 0 âźN(0,I) and solve the ordinary differential equation dZ t dt = v θ (Z t ,t,c), t : 0â 1,(11) to obtain a generated unified representation b T u , which is decoded into an image as b I = D Ď ( b T u ). 3.4IMPLEMENTATION DETAILS Unified tokenizer construction. We instantiate Patch Reparameterization on three pretrained semantic ViTs: SigLIP2-B (Tschannen et al., 2025), DINOv2-B (Oquab et al., 2024), and Qwen-ViT (Bai et al., 2025), obtaining PR-SigLIP2, PR-DINOv2, and PR-Qwen-ViT, respectively. SigLIP2-B is evaluated at 256Ă 256 resolution with a 16 Ă 16 token grid and a 768-dimensional semantic space. DINOv2-B uses its native 14 Ă 14 token grid with a 768-dimensional semantic space. Qwen-ViT supports native-resolution inputs, uses a 16Ă 16 patch size, and has a 1152-dimensional semantic space in our setting. For all backbones, the reconstruction-aware patch embedding P r is initialized from the pretrained patch embedding P s , while all Transformer blocks F Ď are frozen. The reconstruc- tion projection W r maps the reconstruction-aware stream to 128 channels, yielding a 896 unified representation for PR-SigLIP2 and PR-DINOv2 and a 1280 representation for PR-Qwen-ViT. Reconstruction training. For image reconstruction, we use a ViT-XL-scale decoder with 28 layers, hidden width 1152, FFN width 4096, and 16 attention heads. Following the RAE training recipe, the reconstruction objective combines pixel-level and perceptual losses, and the decoder is further refined with an adversarial loss for high-fidelity decoding. For SigLIP2-B and DINOv2-B, we first train P r , W r , and the decoder using L2 and LPIPS losses. We use AdamW with learning rate 2Ă 10 â4 , (β 1 ,β 2 ) = (0.9, 0.95), weight decay 0, global batch size 512, one warmup epoch, and EMA decay 0.9978. We then freeze the encoder-side modules and train only the decoder for another 20 epochs with L2, LPIPS, and GAN losses, using the same optimizer, batch size, and EMA setting. The discriminator 7 UniSpace Table 1: Comparison of visual interfaces in representative MoT-based unified models. âSpacesâ denotes the number of distinct visual representation spaces used by the multimodal model. ModelVisual interfacePretrained semantic prior Separate VAE SpacesReference conditionInterface training BAGEL (Deng et al., 2025)SigLIP2 + FLUX-VAEâ2ViT + clean VAE tokensFrozen SenseNova-U1 (SenseNova-U1 Team, 2026)Native pixelsâ1Clean pixel tokensEnd-to-end UniSpace (Ours)PR-Qwen-ViT tokenizerââ1Unified T u tokensFrozen learning rate follows a cosine schedule from 2Ă10 â4 to 2Ă10 â5 . The main ImageNet results additionally use a short decoder calibration stage: we sample latents from an early DiT checkpoint and update only the decoder, while keeping the encoder and generative model fixed, so that the decoder adapts to realistic generation errors in the unified latent space. SigLIP2-B and DINOv2-B are trained on ImageNet-1K at 256Ă 256 resolution, while Qwen-ViT is trained on web data with the same unified-tokenizer design. Generation Training For ImageNet generation, we evaluate PR-SigLIP2 and PR-DINOv2 following the RAE (Zheng et al., 2025) generation protocol, adopting the same DiTwDDTHead architecture with input size 16, patch size 1, 896 input channels, hidden sizes (1152, 2048), depths (28, 2), and 16 attention heads. We use linear ve- locity prediction with a logit-normal time distribution and the dimension-dependent time-shift rule from RAE (Zheng et al., 2025); for the 896Ă 16Ă 16 latent, the shift uses dimension 229,376 with base 4096. The model is trained with AdamW using learning rate 2Ă 10 â4 , betas (0.9, 0.95), weight decay 0, global batch size 1024 with gradient accumulation 2, EMA decay 0.9995, gradient clipping 1.0, and a linear schedule with 40 warmup epochs decaying to 2Ă 10 â5 by epoch 800. Sampling uses an Euler ODE solver with 50 steps. 4UNISPACE: SCALING UNIFIED MODELING IN ONE VISUAL SPACE Having established a representation that jointly preserves semantic and pixel-level information, we next investigate whether it can serve as the common visual interface of a large-scale multimodal model. We instantiate this idea as UniSpace, a unified multimodal model that performs text-to-image generation, instruction-based image editing, and image understanding in one patch-reparameterized visual representation space. Editing is especially important in this setting because it stress-tests both sides of the representation: the model must understand the reference image and instruction, perform the requested modification, and preserve fine-grained details in all irrelevant regions. UniSpace therefore uses the same frozen unified visual tokenizer for all three tasks and scales it within a Mixture-of-Transformer- Experts architecture. 4.1UNIFIED MOT ARCHITECTURE UniSpace builds on a decoder-only Qwen3-8B (Yang et al., 2025) backbone and adopts the Mixture-of-Transformer- Experts (MoT) design of BAGEL (Deng et al., 2025). It contains an understanding expert and a generation expert that operate on a common multimodal sequence. Token types are hard-routed to modality-specific parameters: text and conditioning-image tokens are processed by the understanding expert, whereas noised visual tokens to be predicted are processed by the generation expert. At every layer, self-attention allows tokens handled by the two experts to interact directly, retaining a bottleneck-free context across understanding and generation. The key distinction of UniSpace lies in its visual interface. Existing MoT-based models typically use two visual representation spaces: a semantic ViT supplies tokens to the understanding pathway, while a separate VAE supplies latents to the generation pathway. In contrast, UniSpace uses the unified encoder and decoder developed in Sec. 3 as its only visual tokenizer, requires neither a separate VAE encoder nor an additional VAE latent space. Table 1 positions this design relative to two representative MoT systems. BAGEL retains two pretrained visual spaces for semantic conditioning and generation, while SenseNova-U1 removes external visual encoders and learns a native pixel interface end-to-end. UniSpace takes a different route: it preserves the semantic prior of a pretrained ViT, reparameterizes it into a reconstruction-compatible visual tokenizer, and consolidates understanding, generation, and editing into one frozen, reusable visual representation without introducing a separate VAE pathway. For image understanding, an input image is encoded as T u and processed together with the text instruction to predict the textual response. For text-to-image generation, the target image is represented in the same T u space, corrupted along the flow path, and predicted by the generation expert conditioned on text. Image editing combines the two cases: the reference image is encoded as T ref u and provided together with the editing instruction, while the noised target representation is predicted in the same T u space. Through shared self-attention, the target tokens condition on 8 UniSpace 1) Understanding Input Image Shared Tokenization and UnifiedEncoding MoTBackbone CELoss Balanced Flow MatchingLoss Whatâs the bird in the image? 2) Text to Image Generation A green bird standing... 3) Image Editing Change the background to sea Text Tokenizer Unified Visual Encoder ... ... ... ... Predicted Text Tokens Predicted Visual Tokens Cross-Entropy / NTP Loss (Understanding) Balanced Flow Matching Loss (Generation & Editing) Instruction Text Prompt Target ImageRef Image Edit Instruction Target Image Figure 3: UniSpace pipeline built on the proposed unified visual representation. UniSpace uses the PR-Qwen-ViT encoderâdecoder as the only visual interface for reference images, target images, and generated images. Text and visual tokens are routed through a Mixture-of-Transformer-Experts backbone for understanding, text-to-image gener- ation, and instruction-based image editing, without an additional VAE pathway. both the instruction and the reference representation, and the sampled representation is decoded by the same decoder D u . Thus, the three tasks differ in token arrangement and routing, but share the same visual tokenizer, representation space, and image decoder, as shown in Fig. 3. 4.2UNIFIED TRAINING OBJECTIVES UniSpace uses two training objectives over a shared multimodal sequence: cross-entropy next-token prediction for textual responses and balanced flow matching for visual prediction in the unified representation space T u . For under- standing samples, the model predicts response tokens with the standard next-token prediction objective L NTP =â L X k=1 logp θ (y k | y <k ,c und ),(12) where y 1:L denotes the target textual response and c und denotes the multimodal context constructed from the input image representation T u and the text instruction. For generation and editing tasks, the target image is first mapped into the unified representation space as Z 1 , and then perturbed following the linear flow path in Eq. 7. We sample the initial noise from a standard Gaussian distribution and define the target velocity as the displacement between the noise and target representation. Following RAE (Zheng et al., 2025), we adopt a logit-normal timestep sampling strategy with the same dimension-dependent time-shift rule. The model is optimized with the balanced component-wise flow matching objective in Eq. 10, where Îť r = 0.75 is used across all UniSpace training stages following the tokenizer-level generation setting. The resulting unified visual prediction objective is L vis (c) = E [â BFM (v θ (Z t ,t,c),V t )].(13) Here, c denotes the task-specific conditioning. For text-to-image generation, the condition contains only text tokens. For image editing, it additionally incorporates the reference image representation and the editing instruction. Thus, L t2i =L vis (c t2i ), L edit =L vis (c edit ).(14) Generation and editing therefore share the same target representation space and visual prediction loss, differing only in their conditioning context. 9 UniSpace Table 2: Progressive training curriculum of UniSpace. T2I, Edit, Und., and VLM denote text-to-image generation, image editing, visual understanding, and visual instruction tuning, respectively. Sample instances refer to the number of examples processed during training rather than the number of unique examples. ItemStage 1Stage 2Stage 3SFT Resolution25651210241024 Training tasksT2I + Und.T2I + Edit + Und.T2I + Edit + Und.T2I + Edit + VLM Sampling ratio10:110:3:110:3:110:3:2 Optimization steps170K115K60K12K Sample instances264.2M189.9M50.9M4.9M Training tokens110B196B151B15B Peak learning rate2Ă 10 â4 1Ă 10 â4 4Ă 10 â5 2Ă 10 â5 Maximum sequence length3,0728,19212,28813,000 Hardware256 NPUs256 NPUs256 NPUs128 NPUs Overall, UniSpace is trained on a mixture of understanding, text-to-image generation, and image-editing samples. Let Ď âund, t2i, edit denote the task type of a training sample. The total training objective can be written as L UniSpace = E (x,Ď)âźD 1 Ď=und L NTP + 1 Ď=t2i L vis (c t2i ) + 1 Ď=edit L vis (c edit ) . (15) Thus, understanding is supervised by cross-entropy over textual responses, while generation and editing share the same balanced flow-matching objective in the unified representation space T u , differing only in their conditioning context. The unified encoder and decoder remain frozen throughout UniSpace training; all MoT parameters are trainable. visual tokenizer remains frozen throughout. 4.3TRAINING DETAILS Training Data and Curriculum. We train UniSpace on internally curated datasets for text-to-image generation, image editing, and visual understanding. These data provide supervision for image generation, instruction-guided editing, and image-text understanding, respectively. The complete training curriculum is summarized in Table 2. We progressively scale training across three stages with increasing image resolutions of 256, 512, and 1024. In Stage 1, we jointly optimize text-to-image generation and visual understanding with a sampling ratio of 10:1. In Stage 2, we introduce image editing and mix generation, editing, and understanding data with a sampling ratio of 10:3:1. Stage 3 further scales all three capabilities through high-resolution packed multimodal training at 1024 resolu- tion. Following pretraining, we perform a multimodal supervised fine-tuning (SFT) stage using packed text-to-image, image-editing, and LLaVA-NeXT-style visual instruction data (Liu et al., 2024b). During SFT, generation, editing, and visual instruction samples are mixed with a sampling ratio of 10:3:2. This stage improves instruction following and visual understanding while preserving the generation and editing capabilities acquired during pretraining. Across all training stages, UniSpace processes approximately 510M training sample instances, corresponding to ap- proximately 470B multimodal tokens. Here, sample instances refer to examples observed during training and may include repeated draws from the underlying datasets. The unified visual tokenizer and visual encoder remain frozen throughout all stages. Optimization. We optimize UniSpace using AdamW with β 1 = 0.9, β 2 = 0.95, Îľ = 10 â15 , and zero weight decay. We use a linear warmup followed by a constant learning-rate schedule. The warmup length is 600 steps for the three pretraining stages and 100 steps for SFT. The peak learning rates for each stage are reported in Table 2. We train in BF16 precision and clip the global gradient norm to 1.0. We use one packed sequence per device at each optimization step without gradient accumulation. No exponential moving average is used during training. The model is trained progressively: each stage is initialized from the selected checkpoint of the preceding stage, while the optimizer and learning-rate scheduler are reinitialized at the beginning of each new stage. The visual encoder and the unified visual tokenizer are frozen throughout training, while the language model, modality experts, and multimodal projection modules are optimized jointly. 10 UniSpace Packed Multimodal Training. To improve training efficiency under heterogeneous multimodal sequence lengths, we employ packed multimodal training. Multiple independent examples, potentially from different training objectives, are concatenated into a single packed sequence until a predefined token budget is reached. Samples are drawn accord- ing to the task-specific sampling ratios in Table 2. Consequently, the specified ratios describe the expected sampling frequencies over the full training run rather than the exact composition of every individual packed sequence. We construct block-wise attention masks for packed sequences to prevent information exchange between different examples. In particular, tokens may only attend to tokens belonging to the same example, while attention within each example follows its corresponding modality-specific pattern. Text tokens use causal attention, whereas visual tokens use full attention within the same sample. Position indices are reset for every individual example inside a packed sequence. Losses are computed only over valid target tokens: the text cross-entropy loss is applied to target text tokens, while the image generation loss is applied only to target visual tokens. The maximum packed sequence lengths are 3,072, 8,192, 12,288, and 13,000 for Stage 1, Stage 2, Stage 3, and SFT, respectively. The resulting token utilization is approximately 73% to 79% across stages. As the image resolution increases, each packed sequence contains fewer examples due to the increased number of visual tokens; however, the global task sampling distribution remains unchanged. Data Processing. We use resolution-specific image buckets for the three pretraining stages. Images are resized and bucketed according to their aspect ratios, with an aspect-ratio deviation threshold of 0.05. Visual inputs are tokenized using the frozen unified visual tokenizer. For image generation, we apply text conditioning dropout with probability 0.1 to support classifier-free guidance. We also apply instruction dropout with probability 0.1 during multimodal training. For image editing, reference images are processed through the understanding branch, while the target image is represented as a generation target. Distributed Training Infrastructure. We train Stages 1â3 on 256 Ascend 910B NPUs with 64GB memory each, distributed over 16 nodes with 16 NPUs per node. The SFT stage is trained on 128 Ascend 910B NPUs over 8 nodes. Training is implemented in PyTorch using FSDP with the SHARD GRADOP sharding strategy. We use BF16 mixed precision, activation checkpointing, and SDPA-based attention to reduce memory consumption. The selected training trajectory requires approximately 142K NPU-hours in total. 5EXPERIMENTS 5.1UNIFIED TOKENIZER EVALUATION For reconstruction evaluation, we use the ImageNet-1K validation set to assess PR-SigLIP2 and PR-DINOv2, report- ing PSNR, rFID, and LPIPS as evaluation metrics. For multimodal understanding, we follow the LLaVA-v1.5 (Liu et al., 2024a) setup, using Vicuna-7B-v1.5 as the language model and a two-layer MLP as the multimodal projector. The visual encoder is frozen, and the complete unified representation T u is provided as input to the projector. For ImageNet generation, we evaluate our method on ImageNet using rFID. We report results from checkpoints trained for 80 and 800 epochs. 5.1.1HIGH-FIDELITY IMAGE RECONSTRUCTION We first assess whether Patch Reparameterization equips pretrained semantic encoders with high-fidelity image re- construction capability. As shown in Table 3, the proposed method consistently achieves strong reconstruction quality across PR-SigLIP2, PR-DINOv2, and PR-Qwen-ViT. In particular, PR-DINOv2 attains an rFID of 0.14, a PSNR of 30.84, and an SSIM of 0.90, achieving the best rFID among the compared tokenizers with demonstrated semantic capability. Figure 4 presents qualitative comparisons on representative ImageNet validation images, with red boxes highlighting fine local details and text-like regions that are particularly sensitive to reconstruction quality. The matched-backbone comparisons further demonstrate the effectiveness of Patch Reparameterization. Compared with RAE using the corresponding SigLIP2-B and DINOv2-B encoders, PR-SigLIP2 reduces rFID from 0.53 to 0.18, while PR-DINOv2 reduces it from 0.57 to 0.14, corresponding to relative reductions of 66.0% and 75.4%, respectively. For PR-DINOv2, PSNR increases from 18.86 to 30.84 and SSIM from 0.48 to 0.90. Our method also compares favorably with the recent RAEv2 reconstruction-enhanced variant: PR-DINOv2 improves PSNR from 22.57 to 30.84 and rFID from 0.29 to 0.14, while using the smaller DINOv2-B encoder rather than the DINOv3-L backbone used by RAEv2. These improvements are achieved with the pretrained ViT backbone kept frozen, indicating that its Transformer blocks can effectively propagate reconstruction-relevant visual details when driven by a reconstruction- aware patch embedding. 11 UniSpace Input ImageOursRAEVA-VAEVTP Figure 4: Qualitative reconstruction comparison. Columns show the input image, Ours, RAE, VA-VAE, and VTP. Red boxes highlight local details and text-like regions. 12 UniSpace Table 3: Reconstruction quality on the ImageNet-1K 256Ă 256 validation set. âRatioâ denotes the spatial downsam- pling ratio. MethodTypeTraining DataRatio PSNRâ SSIMâ rFIDâ Tokenizers without Demonstrated Semantic Capability LlamaGen (Sun et al., 2024)Discrete-PixelMS+IN-1K1620.650.542.47 Open-MAGVIT2 (Luo et al., 2024)Discrete-PixelMixed-100M1622.700.641.67 SD-VAE XL (Rombach et al., 2021)Continuous-PixelOImg+LAae++827.370.780.67 Qwen-Image (Wu et al., 2025a)Continuous-Pixelâ832.180.901.45 SD-VAE 3 (Rombach et al., 2021)Continuous-Pixelâ831.290.870.20 Wan2.1 (Wan et al., 2025)Continuous-Pixelâ831.340.890.95 FLUX-VAE (Labs, 2024)Continuous-Pixelâ832.740.920.18 VA-VAE (Yao et al., 2025b)Continuous-PixelIN-1K1627.960.790.28 Wan2.2 (Wan et al., 2025)Continuous-Pixelâ1631.250.880.74 Tokenizers with Demonstrated Semantic Capability Show-o (Xie et al., 2025)Discrete-Pixelâ1621.340.593.50 QLIP-B (Zhao et al., 2025)Discrete-PixelDC-1B1623.160.633.21 VILA-U (Wu et al., 2025c)Discrete-PixelWL-10B+CY-1B16â1.80 TokenFlow (Qu et al., 2025)Discrete-PixelLA+CY1621.410.691.37 UniTok (Ma et al., 2026)Discrete-PixelDC-1B1627.280.770.41 UniLIP (Tang et al., 2025)Continuous-PixelBP-32M3222.990.750.79 BLIP3-o (Chen et al., 2025a)Continuous-DiffusionBP-32M1614.710.583.18 VTP-L (Yao et al., 2025a)Continuous-PixelDC-277M1625.820.740.36 UniFlow (SigLIP2) (Yue et al., 2025)Continuous-DiffusionIN-1K1629.380.930.62 UniFlow (DINOv2) (Yue et al., 2025)Continuous-DiffusionIN-1K1431.010.940.54 RAE (SigLIP2-B) (Zheng et al., 2025)Continuous-PixelIN-1K1619.350.490.53 RAE (DINOv2-B) (Zheng et al., 2025)Continuous-PixelIN-1K1418.860.480.57 RAEv2 (DINOv3-L, K=7) â (Singh et al., 2026)Continuous-PixelIN-1K1622.570.630.29 PR-SigLIP2Continuous-PixelIN-1K1629.640.870.18 PR-DINOv2Continuous-PixelIN-1K1430.840.900.14 PR-Qwen-ViTContinuous-PixelWeb Data1630.160.880.17 Note. â For RAEv2, we report the K=7 variant, which is used as its reconstructionâgeneration trade-off setting. VTP-L reports semantic capability through ImageNet zero-shot and linear probing evaluation. The proposed representations are also competitive with tokenizers specifically designed for pixel reconstruction and generation. PR-DINOv2 improves over VA-VAE by reducing rFID from 0.28 to 0.14, and even surpasses large-scale generative VAEs such as FLUX-VAE and SD-VAE 3 in terms of rFID. Meanwhile, PR-Qwen-ViT shows that Patch Reparameterization generalizes beyond ImageNet-pretrained encoders. Crucially, unlike reconstruction-specialized tokenizers, our representations retain the semantic capabilities of their respective pretrained encoders, as evaluated in the next section. 5.1.2MULTIMODAL UNDERSTANDING We next examine whether Patch Reparameterization compromises the semantic understanding inherited from the pre- trained encoder. We compare against the corresponding original semantic encoders under the same LLaVA-v1.5 setup, using identical language models, multimodal projectors, training data, and evaluation protocols. In all experiments, the complete unified representation T u , rather than the semantic component T s alone, is fed to the multimodal projector. As shown in Table 4, Patch Reparameterization does not compromise the overall multimodal understanding capability of the original encoders. PR-SigLIP2 achieves an average score of 64.37, surpassing the original SigLIP2-B baseline of 63.39, while PR-Qwen-ViT reaches 68.94, slightly above the original Qwen-ViT baseline of 68.29. Beyond the matched-backbone comparisons, our method also achieves highly competitive understanding performance under the LLaVA-v1.5/Vicuna-7B setting. These results demonstrate that Patch Reparameterization retains the semantic capa- bility of pretrained encoders while incorporating reconstruction-aware visual details into the unified representation. 13 UniSpace Table 4: Multimodal understanding performance of unified visual tokenizers. Our patch-reparameterized tokenizers are evaluated using the complete unified representation T u . Methods marked withâ use the LLaVA-v1.5 (Liu et al., 2024a) training data. The âOriginal Semantic Encodersâ rows are matched baselines using the same semantic encoder, LLM, training data, and evaluation protocol as our corresponding variants. MethodVisual EncoderLLMRes. POPE GQA TQA MMV MMB MME-S MME-P Avg. Existing Unified Visual Tokenizers VILA-Uâ (Wu et al., 2025c)SigLIP-SO400MVicuna-7B25681.6â 1311.6â UniTokâ (Ma et al., 2026)ViTamin-LVicuna-7B25681.7â 1448.0â QLIPâ (Zhao et al., 2025)CLIP-LVicuna-7B39286.1 61.8 55.233.3â 1498.3â TokenFlow-Bâ (Qu et al., 2025)CLIP-BVicuna-13B22484.0 59.3 49.822.455.3 1660.4 1353.6 60.21 TokenFlow-Lâ (Qu et al., 2025)ViTamin-XLVicuna-13B25685.0 60.3 54.127.760.3 1622.9 1365.4 62.40 UniTok (Ma et al., 2026)ViTamin-LLLaMA-2-7B 25683.2 61.1 51.633.9â 1448.0â TokLIP (Lin et al., 2025b)VQ-GAN+ViT-SO400MQwen2.5-7B 38484.1 59.5â29.867.6â 1448.4â TokenFlow-XL (Qu et al., 2025)SigLIP-SO400MQwen2.5-14B 38487.8 62.5 62.348.276.8 1922.2 1551.1 73.04 UniFlow-LVâ (Yue et al., 2025)DFN-CLIP-LVicuna-7B224 86.56 61.38 53.4030.2 63.83 1748.0 1446.9 65.02 UniFlow-LVâ (Yue et al., 2025)SigLIP2-SO400MVicuna-7B256 87.94 63.29 58.032.4 68.38 1823.0 1477.9 67.87 UniFlow-LVâ (Yue et al., 2025)DINOv2-LVicuna-7B378 88.04 59.37 45.5325.6 51.48 1590.5 1257.7 58.92 UniFlow-LVâ (Yue et al., 2025) InternViT-300M (Chen et al., 2024)Vicuna-7B448 88.97 63.35 61.8536.6 67.10 1803.0 1505.1 69.04 Original Semantic Encoders SigLIP2 baselineâ SigLIP2-BVicuna-7B25685.2 61.4 54.3625.064.4 1689.3 1378.2 63.39 Qwen-ViT baselineâ Qwen-ViTVicuna-7B448 86.26 63.25 64.27 31.88 69.24 1791.3 1471.3 68.29 Ours: Patch-Reparameterized Unified Visual Tokenizers PR-SigLIP2â PR-SigLIP2Vicuna-7B25684.8 61.1 54.625.565.2 1741.1 1447.2 64.37 PR-Qwen-ViTâ PR-Qwen-ViTVicuna-7B448 86.57 63.13 63.68 30.28 69.07 1874.5 1522.7 68.94 Note: The average is computed as Avg = (POPE + GQA + TQA + MMV + MMB + MME-S/20 + MME-P/20)/7. Table 5: System-level reconstructionâgeneration performance on ImageNet 256Ă 256 with representative latent gen- erative priors, following the evaluation protocols reported by VA-VAE (Yao et al., 2025b) and RAE (Zheng et al., 2025). Reconstruction ModelGeneration w/o CFGGeneration w/ CFG MethodTokenizerrFIDâ#Params gFIDâ sFIDâ ISâ Pre.â Rec.â gFIDâ sFIDâ ISâ Pre.â Rec.â Representative Latent Generative Priors REPA (Yu et al., 2024)SD-VAE0.61675M5.90â1.424.70 305.7 0.80 0.65 LightningDiT (Yao et al., 2025b)VA-VAE0.28675M2.174.36 205.6 0.77 0.651.354.15 295.3 0.79 0.65 UniFlow (MAR) (Yue et al., 2025)UniFlow (InternViT)0.28479M2.45â228.0â1.85â290.0â VTP-L (LightningDiT) (Yao et al., 2025a)VTP-L0.36675M1.85â232.3 0.79 0.631.11â279.5 0.79 0.67 RAE (DiT DH -XL) (Zheng et al., 2025)RAE (DINOv2-B)0.57839M1.51â242.9 0.79 0.631.13â262.6 0.78 0.67 RAEv2 (DiT DH -XL) (Singh et al., 2026)RAEv2 (DINOv3-L, K=7)0.290.9B1.65â228.0â1.06â255.3â Ours: Patch-Reparameterized Unified Visual Tokenizers PR-SigLIP2PR-SigLIP20.18839M4.426.66 190.4 0.72 0.652.806.05 248.2 0.78 0.61 PR-DINOv2PR-DINOv20.14839M2.105.39 217.2 0.78 0.641.874.89 274.3 0.82 0.60 Note. For RAEv2, we report the K=7 reconstructionâgeneration trade-off setting from the original paper. âââ indicates metrics not reported in the corresponding source under the matched setting. 5.1.3IMAGENET GENERATION More importantly, our method establishes a favorable reconstructionâgeneration operating point for latent generative modeling. As shown in Table 5, PR-DINOv2 shifts the trade-off toward substantially higher reconstruction fidelity while retaining effective generation performance, achieving an rFID of 0.14 and gFIDs of 2.10 without classifier-free guidance and 1.87 with classifier-free guidance at scale 1.2. In contrast, RAE, RAEv2, and VTP-L achieve stronger generation results but have substantially higher tokenizer rFID values of 0.57, 0.29, and 0.36, respectively. Thus, Patch Reparameterization recovers substantially more decoder-relevant visual detail from a pretrained semantic ViT while still producing a latent space that can be effectively modeled by a generative prior. Taken together with the reconstruction and multimodal understanding results, these findings demonstrate that the pro- posed representation is not merely a reconstruction tokenizer or a generation latent. Instead, it provides a unified visual 14 UniSpace 1) Understanding Whatâs the bird 2) Text to Image Generation 3) Image Editing Change the Semantic Encoder í¸ ! High-Fidelity Reconstruction Decoder íˇ " RAE-style Semantic Decoder íˇ ! Reconstruction Encoder í¸ " DIT Trained on Merged Unified Token Semantic DIT Reconstruction + understanding â generation-friendly representation Semantic-Pixel Decoder PSNR:33.83 rFID:.069 Zeroshot Acc:78.53 gFID 8.07 gFID 120.9 Strong Representation in Reconstrution and Semantic Trained on frozen semantic latents Trained for high-fidelity reconstruction Step1: Separate Decoder Training Semantic Representation Reconstruction Representation í " Concat+MLP Merger M Semantic í ! Reconstruction í " Strong representation-level metrics Zero-shot Acc: 78.53 SigLIP baseline: 79.10 PSNR: 33.83 rFID:0.069 Step2: Entangled Unified Representation Entangled Unified Token í # â˘Cosine align to í ! â˘Linear align to í " Step3: Generative Modeling on Zm DiT trained on entangled tokens 20 epochs Generated latent " í # í ! Decode with íˇ ! FID: 8.07 Decode with íˇ " FID: 120.9 Takeaway Semantic directions are learned, but reconstruction-critical directions are missed. Entangled tokens are dominated by semantic variation and are difficult to decode faithfully with the high-fidelity reconstruction decoder. Figure 5: Diagnostic study of an entangled unified representation. We construct an MLP-merged token Z m = M(Z s ,Z r ) from semantic and reconstruction representations and align it with both. Although Z m supports strong understanding and high-fidelity reconstruction on real encoded latents, its generative decoding fails when using the pretrained high-fidelity reconstruction decoder D r , which yields an FID of 120.9. Decoding the same generated latent with the semantic decoder D s gives a much lower FID of 8.07, indicating that the generative prior captures semantic structure but fails to preserve reconstruction-relevant information in the entangled latent space. representation that retains multimodal understanding, supports high-fidelity reconstruction, and remains effective for latent image generation. 5.1.4AUXILIARY DIAGNOSTIC STUDY ON ENTANGLED UNIFIED REPRESENTATIONS As an auxiliary diagnostic experiment, separate from our final design, we investigate whether jointly supporting se- mantic understanding and high-fidelity reconstruction is alone sufficient for generative modeling. We construct a deliberately entangled representation by merging semantic and reconstruction features into a single latent space with- out explicit decomposition. As shown in Fig. 5, this diagnostic variant performs well when evaluated on real encoded latents, achieving a zero-shot accuracy of 78.53 (compared with 79.10 for the SigLIP baseline), an rFID of 0.069, and a PSNR of 33.83. However, the entangled representation fails to support faithful generative decoding. After training a DiT prior on the merged representation, decoding the generated latent b Z m with the pretrained high-fidelity reconstruction decoder D r results in an FID of 120.9, indicating a severe generation failure. Interestingly, decoding the same b Z m with the semantic decoder D s yields a much lower FID of 8.07. This result does not indicate successful generation in the unified latent space; rather, it shows that the generative prior mainly captures the dominant semantic structure, while failing to preserve the reconstruction-relevant information required by D r . We further find that approximately 95% of the merger output variation is explained by the semantic pathway. This semantic dominance explains the decoder-dependent generation gap: although the generated latent retains enough semantic structure for D s to produce plausible images, it does not preserve the fine-grained directions needed for high- fidelity decoding by D r . These results reveal a gap between representation quality and generative modelability: good reconstruction on real encoded latents does not guarantee faithful reconstruction after generation. This motivates our explicit factorization, in which T s and e T r are concatenated along the channel dimension, making the reconstruction component directly accessible during flow-matching training. Further details of the entangled representation and diagnostic protocol are provided in Appendix C. 15 UniSpace Table 6: Quantitative comparison on ImgEdit. All category scores and the overall score are higher-is-better. Baseline values are reproduced from the evaluation reported by SenseNova-U1 (SenseNova-U1 Team, 2026). âActiveâ denotes activated generation parameters for sparse models. Model# Params Add Adjust Extract Replace Remove Bg. Style Hybrid Action Overallâ Closed-source Models UniWorld-V2 (Li et al., 2025)â4.294.444.324.694.724.414.913.834.834.49 Nano-Banana-Proâ4.444.62 3.424.604.634.32 4.973.644.694.37 Seedream 4.5â4.574.652.974.664.464.37 4.923.714.564.32 Seedream 4.0â4.334.383.894.654.574.35 4.223.714.614.30 Nano-Bananaâ4.624.413.684.344.394.40 4.183.724.83 4.29 GPT-Image-1â4.614.332.904.353.664.57 4.933.964.894.20 FLUX.1 Kontext [Pro] (Labs et al., 2025)â4.254.152.354.563.574.26 4.573.684.634.00 Open-source Image Generation and Editing Models Qwen-Image-Edit-2511 (Wu et al., 2025a)20B4.544.574.134.704.464.36 4.894.164.814.51 LongCat-Image-Edit (Team et al., 2025)6B4.444.53 3.834.804.604.33 4.923.754.824.45 FLUX.2 [Dev] (Labs, 2024)32B4.504.183.834.654.654.31 4.883.464.704.35 Qwen-Image-Edit-2509 (Wu et al., 2025a)20B4.324.364.044.644.524.374.843.394.714.35 Z-Image-Edit (Cai et al., 2025)6B4.404.144.304.574.134.14 4.853.634.504.30 Qwen-Image-Edit (Wu et al., 2025a)20B4.384.163.434.664.144.38 4.813.824.694.27 FLUX.1 Kontext [Dev] (Labs et al., 2025)12B4.123.802.044.223.093.97 4.513.354.253.71 OmniGen2 (Wu et al., 2025b)4B3.573.061.773.743.203.57 4.812.524.683.44 Step1X-Edit (Liu et al., 2025)12B3.883.141.763.402.413.16 4.632.642.523.06 OmniGen (Xiao et al., 2025)3.8B3.473.041.712.942.433.21 4.192.243.382.96 Open-source Unified Multimodal Models Emu3.5 (Cui et al., 2025)32B4.614.323.964.844.584.35 4.793.694.574.41 Ovis-U1 (Wang et al., 2025)1.2B3.993.732.664.384.154.05 4.863.43 4.683.97 SenseNova-U1 (SenseNova-U1 Team, 2026)8B3.834.153.124.323.264.18 4.85 3.034.413.90 InternVL-U (w/ CoT) (Tian et al., 2026)1.7B4.243.802.584.363.513.92 4.693.004.313.82 InternVL-U (Tian et al., 2026)1.7B4.133.402.274.133.393.84 4.773.034.053.67 UniWorld-V1 (Lin et al., 2025a)12B3.823.642.273.473.242.99 4.212.962.743.26 BAGEL (Deng et al., 2025)7B3.563.311.703.302.623.24 4.492.384.173.20 UniSpace (Ours)8B4.534.383.614.674.424.234.552.704.474.28 5.2SCALING UNIFIED MULTIMODAL MODELING We next study whether the proposed unified representation can be scaled from controlled representation learning to a full multimodal system. UniSpace is not intended to isolate the representation alone: it combines the frozen unified encoderâdecoder with an 8B Mixture-of-Transformer-Experts backbone and is trained on large-scale internal generation, editing, and understanding data. We therefore use UniSpace as a system-level validation. The central question is whether a single ViT-based visual space can replace the conventional separation between semantic image tokens and VAE latents while still supporting practical generation and editing. 5.2.1IMAGE EDITING We evaluate UniSpace on two image-editing benchmarks, ImgEdit (Ye et al., 2025) and GEdit (SenseNova-U1 Team, 2026). ImgEdit serves as our primary benchmark for system-level comparison, while GEdit provides a complementary evaluation of instruction following and perceptual quality under both English and Chinese prompts. ImgEdit. ImgEdit covers nine editing categories, ranging from local content manipulation to style transfer and hy- brid editing. As shown in Table 6, UniSpace achieves an overall score of 4.28 at the 8B scale. It substantially outper- forms comparable-scale unified models, including SenseNova-U1 (3.90) and BAGEL (3.20), while approaching the 4.41 score of the substantially larger 32B Emu3.5. These results demonstrate the strong editing capability of UniSpace at the 8B scale. Despite using substantially fewer parameters than Emu3.5, UniSpace achieves a comparable overall score, while clearly outperforming other unified models of similar scale. This suggests that the unified representation provides sufficient semantic and reconstruction information for precise instruction-based image editing. 16 UniSpace Table 7: Additional image-editing evaluation on GEdit. Reference GEdit-Bench-EN results are reproduced from SenseNova-U1 (SenseNova-U1 Team, 2026); BAGEL-CN is reproduced from the BAGEL (Deng et al., 2025) report; UniSpace is evaluated with GPT-4o. Avg. is computed over English and Chinese overall scores when both are available. All metrics are higher-is-better. GEdit-Bench-ENGEdit-Bench-CN Model# Params SCPQ Overallâ SCPQ Overallâ Avg.â Specialized Image Editing Models Qwen-Image-Edit-2511 (Wu et al., 2025a)20B8.30 8.207.88â LongCat-Image-Edit (Team et al., 2025)6B8.138.187.75â Z-Image-Edit (Cai et al., 2025)6B8.11 7.727.57â Qwen-Image-Edit (Wu et al., 2025a)20B8.00 7.867.56â Unified Multimodal Models Emu3.5 (Cui et al., 2025)32B8.11 7.707.59â SenseNova-U1 (SenseNova-U1 Team, 2026)8B8.277.497.47â BAGEL (Deng et al., 2025)7B7.36 6.836.527.346.856.506.51 UniSpace (Ours)8B8.29 7.067.418.27 7.007.387.395 GEdit. We further evaluate UniSpace on GEdit. As GEdit serves as a complementary benchmark, we report aggre- gate results and compare UniSpace with other representative methods in Table 7. UniSpace achieves overall scores of 7.41 and 7.38 on the English and Chinese full sets, respectively, yielding a bilingual average of 7.39. It outperforms BAGEL and remains close to the dense SenseNova-U1 8B model. The detailed results show that UniSpace achieves strong semantic consistency, while its perceptual quality remains relatively weaker, particularly on compositional editing and other perceptually sensitive cases. Qualitative Results. Figure 6 presents qualitative comparisons on four representative editing tasks: object addition, style transfer, scene replacement, and object substitution. UniSpace consistently follows the editing instructions while preserving the identity and structure of unrelated image regions. In contrast, BAGEL and SenseNova-U1 tend to produce artifacts or exhibit incomplete instruction adherence in challenging cases, particularly for fine-grained style transfer and multi-object compositional editing. Figure 7 shows additional results on human-centric editing tasks, including expression change, accessory addition, clothing color modification, and hairstyle transfer. These examples require fine-grained localized editing while preserving identity and unrelated appearance attributes. UniSpace per- forms these edits while maintaining the overall appearance and structural consistency of the input subjects, further supporting its ability to combine instruction following with localized visual editing. 5.2.2TEXT-TO-IMAGE GENERATION We evaluate UniSpace on three text-to-image generation benchmarks: GenEval (Ghosh et al., 2023), OneIG- Bench (Chang et al., 2025), and DPG-Bench (Hu et al., 2024). GenEval evaluates object-centric compositional alignment, OneIG-Bench measures fine-grained generation across English and Chinese prompts, and DPG-Bench evaluates dense-prompt following across global, entity, attribute, and relational constraints. We adopt the comparison sets collected by SenseNova-U1 (SenseNova-U1 Team, 2026) and report the results in Tables 8â11. GenEval. As shown in Table 8, UniSpace achieves an overall score of 0.84. UniSpace performs well on single-object generation, two-object generation, positional alignment, and attribute binding, although its counting and color-binding scores remain relatively weaker. OneIG-Bench. On OneIG-Bench, UniSpace obtains overall scores of 0.561 and 0.533 on the English and Chinese subsets, respectively, yielding a bilingual average of 0.547. As shown in Tables 9 and 10, UniSpace achieves the strongest bilingual average among the compared unified multimodal models, slightly exceeding Emu3.5 (0.546) and SenseNova-U1 (0.542). It also obtains the highest style scores among the compared methods on both the English and Chinese subsets, with scores of 0.467 and 0.455, respectively. These results demonstrate that UniSpace remains competitive in multilingual text-to-image generation, particularly in style control and fine-grained visual synthesis. 17 UniSpace Change the construction site in the picture to a beach scene. Replace the butterfly in the image with a squirrel. Transfer the image into a Lego-brick stop-motion diorama style. Add a red sports car in the center of the image, positioned facing towardsthe right. Input Image Ours Bagel SenseNova-U1 Figure 6: Qualitative image editing comparison. Columns (left to right): input image, Ours (UniSpace), BAGEL, SenseNova-U1. Rows (top to bottom): object addition, style transfer, background replacement, object replacement. 18 UniSpace Change the person's blue graduation robes to green graduation robes. Change the woman's long braided hair to loose, voluminous curly hair. Add a black graduation cap to the person's head. Change the woman's neutral expression to a cheerful smile. Input ImageOursBagelSenseNova-U1 Change the woman's smiling expression to a serious one. Figure 7: Qualitative comparison on human-centric image editing. Columns (left to right): input image, Ours (UniS- pace), BAGEL, SenseNova-U1. Rows: expression change (neutralâsmile), accessory addition (graduation cap), clothing recoloring (blueâgreen robes), hairstyle transfer (braidsâcurly), expression change (smileâserious). 19 UniSpace Table 8: Text-to-image generation on GenEval. SO, TO, CT, CL, POS, and ATTR denote single object, two objects, counting, colors, position, and attribute binding, respectively. All metrics are higher-is-better. Model# ParamsSOTOCTCL POS ATTR Overallâ Open-source Generation Models Qwen-Image (Wu et al., 2025a)20B0.990.920.89 0.88 0.760.770.87 Z-Image (Cai et al., 2025)6B1.00 0.94 0.780.93 0.620.770.84 SD3-Medium (Rombach et al., 2021)2B0.990.94 0.72 0.890.330.600.74 FLUX.1-dev (Labs, 2024)12B0.98 0.81 0.74 0.79 0.220.450.66 Open-source Unified Multimodal Models SenseNova-U1 (SenseNova-U1 Team, 2026)8B1.00 0.960.92 0.92 0.910.760.91 InternVL-U (Tian et al., 2026)1.7B0.990.94 0.74 0.910.770.740.85 BAGEL (Deng et al., 2025)7B0.99 0.94 0.81 0.88 0.640.630.82 Janus-Pro (Chen et al., 2025b)7B0.990.89 0.59 0.90 0.790.660.80 OmniGen2 (Wu et al., 2025b)4B1.00 0.95 0.64 0.88 0.550.760.80 UniWorld-V1 (Lin et al., 2025a)12B0.990.93 0.79 0.89 0.490.700.80 Show-o2 (Xie et al., 2026)7B1.00 0.87 0.58 0.92 0.520.620.76 Emu3.5 (Cui et al., 2025)32Bâ0.73 UniSpace (Ours)8B0.98 0.92 0.69 0.88 0.830.730.84 DPG-Bench. On DPG-Bench, UniSpace achieves an overall score of 86.49. It outperforms BAGEL (85.07), Show- o2 (86.14), and most other listed unified multimodal models. UniSpace achieves the highest relation score among all compared models, reaching 94.97, and obtains a strong entity score of 92.26. These results indicate that UniSpace can effectively follow dense prompts, particularly those involving relational and entity-level constraints, although its global and attribute scores remain below those of the strongest text-to-image systems. Qualitative Results. Figures 8 and 9 present qualitative text-to-image comparisons between UniSpace, BAGEL, and SenseNova-U1. UniSpace generally produces images that are well aligned with the input prompts while maintaining coherent object layouts, attributes, and visual styles. Compared with the other unified models, its generations exhibit a more realistic appearance and richer visual details, especially in complex scenes and fine-grained visual content. These results provide qualitative evidence that the unified representation supports both prompt alignment and high-fidelity visual synthesis. 5.2.3UNIFIED CAPABILITY EVALUATION The same model also retains an understanding pathway, but we do not use UniSpaceâs system-level understanding scores as the main evidence for this paper. UniSpace is optimized primarily for generation and editing, and its post- training recipe is not designed to maximize standard VLM benchmarks. Instead, the understanding capability of the unified representation itself is validated in the controlled experiments in Table 4, where all tasks use the complete unified representation T u . 6LIMITATION Our results establish the viability of a shared visual representation, but do not imply uniform improvements across all downstream tasks. UniSpace is primarily optimized for generation and editing, and its system-level understanding per- formance still lags behind that of dedicated visionâlanguage models, despite controlled experiments showing that the representation itself largely preserves the capabilities of the original semantic encoder. Moreover, because the encoder serves as a unified representation for both understanding and generation, it cannot be freely updated during training on understanding data without potentially compromising its generative capabilities. This architectural constraint limits further improvements in visual understanding. An important direction for future work is therefore to explore how the 20 UniSpace OursBagelSenseNova-U1 In a dense, tropical jungle, a vibrant green and yellow parrot is perched elegantly on a twisted branch. The birdâs plumage glows under dappled sunlight ...... An elegant portrait of a Siberian tiger prowling through a snowy forest landscape. The tiger's striking orange and black stripes stand out boldly against the pristine ...... A meticulously styled seafood platter featuring an array of freshly shucked oysters on a bed of crushed ice, dotted with lemon wedges and sprigs of dill. The oysters ...... A close-up of a rare black rose in a garden, capturing the velvety texture and deep color of its petals. Water droplets, remnants of a recent rain shower, cling to the petals, ...... Figure 8: Qualitative text-to-image comparison. Columns show Ours (UniSpace), BAGEL, and SenseNova-U1. Each row is generated from the prompt shown below the corresponding images. 21 UniSpace OursBagelSenseNova-U1 A close-up portrait of an elderly woman with deep-set wrinkles and kind eyes, her expression serene. Her silver hair is tied in a loose bun, wisps catching the soft ...... A serene portrait of a young woman meditating by a tranquil lake at dawn. Her posture is relaxed, eyes closed, with morning mist gently enveloping her surroundings. ...... A stark and minimalist photograph of a single, barren tree in the midst of a snow-covered landscape. The contrast between the dark silhouette of the tree and the pristine,...... A vivid, sunlit image of a blooming cherry blossom tree in full spring glory. The myriad of pink and white blossoms create a cloud-like effect, set against a bright, ...... Figure 9: Additional qualitative text-to-image comparison. Columns show Ours (UniSpace), BAGEL, and SenseNova- U1. Each row is generated from the prompt shown below the corresponding images. 22 UniSpace Table 9: Text-to-image generation on OneIG-Bench (English). Model# Params Align.Text Reason.StyleDiv. Overallâ Closed-source Models Gemini-2.5-Flash-Imageâ0.878 0.8940.346 0.4500.1820.550 GPT-Image-1â0.8510.8570.3450.462 0.1510.533 Seedream 3.0â0.818 0.8650.275 0.413 0.2770.530 Open-source Generation Models Qwen-Image (Wu et al., 2025a)20B0.882 0.8910.306 0.418 0.1970.539 Open-source Unified Multimodal Models Emu3.5 (Cui et al., 2025)32B0.902 0.9940.345 0.427 0.1510.564 SenseNova-U1 (SenseNova-U1 Team, 2026)8B0.8820.9690.3300.396 0.1660.549 BAGEL (Deng et al., 2025)7B0.769 0.2440.173 0.367 0.251 0.361 Janus-Pro (Chen et al., 2025b)7B0.553 0.0010.139 0.276 0.3650.267 UniSpace (Ours)8B0.860 0.9370.311 0.467 0.2330.561 Table 10: Text-to-image generation on OneIG-Bench (Chinese). Baseline results are reproduced from SenseNova- U1 (SenseNova-U1 Team, 2026). All metrics are higher-is-better. Model# Params Align.Text Reason.StyleDiv. Overallâ Closed-source Models Gemini-2.5-Flash-Imageâ0.825 0.2760.2980.4270.1980.337 GPT-Image-1â0.8120.6500.300 0.449 0.1590.474 Seedream 3.0â0.793 0.9280.281 0.397 0.2430.528 Open-source Generation Models Qwen-Image (Wu et al., 2025a)20B0.825 0.9630.267 0.405 0.2790.548 Open-source Unified Multimodal Models Emu3.5 (Cui et al., 2025)32B0.853 0.9410.3000.386 0.1660.529 SenseNova-U1 (SenseNova-U1 Team, 2026)8B0.8260.9770.303 0.3920.1760.535 BAGEL (Deng et al., 2025)7B0.672 0.3650.186 0.357 0.268 0.370 Janus-Pro (Chen et al., 2025b)7B0.324 0.1480.104 0.264 0.3580.240 UniSpace (Ours)8B0.807 0.8810.276 0.455 0.2440.533 unified representation and UniMM can be jointly optimized while maintaining a balance between understanding and generation. 7CONCLUSION We presented Patch Reparameterization, a minimal adaptation that turns a pretrained semantic ViT into a unified visual tokenizer without modifying its Transformer blocks. By retaining the original semantic pathway, introducing a reconstruction-aware patch pathway, and explicitly factorizing their outputs, the proposed representation supports multimodal understanding, high-fidelity reconstruction, and generation within a single frozen parameter space. More broadly, our findings suggest that unifying visual capabilities need not begin with training a new vision backbone: changing how information enters an existing semantic encoder can be sufficient to expose a substantially broader range of visual information. Scaling the same frozen representation to UniSpace further shows that a single visual space can replace the con- ventional ViTâVAE dual interface in a large multimodal system. Within an 8B mixture-of-experts model, UniSpace 23 UniSpace Table 11: Dense-prompt following performance on DPG-Bench. All metrics are higher-is-better. Model# Params Global Entity Attribute Relation Other Overallâ Closed-source Models Seedream 4.5â89.2494.3092.1492.2393.8388.63 Nano-Banana-Proâ91.0092.85 91.5692.3989.9387.16 GPT-Image-1â88.8988.9489.8492.6390.9685.15 Open-source Generation Models Qwen-Image (Wu et al., 2025a)20B91.3291.5692.0294.3192.7388.32 Z-Image (Cai et al., 2025)6B93.3991.2293.1692.2291.5288.14 SD3-Medium (Rombach et al., 2021)2B87.9091.0188.8380.7088.6884.08 FLUX.1-dev (Labs, 2024)12B74.3590.0088.9690.8788.3383.84 Open-source Unified Multimodal Models SenseNova-U1 (SenseNova-U1 Team, 2026)8B88.7490.9092.4392.4392.5087.78 Tuna (Liu et al., 2026)7B90.42 91.6890.9491.8790.7386.76 NEO-unify (SenseNova-U1 Team, 2026)8B91.0091.5392.0694.1490.4386.71 Show-o2 (Xie et al., 2026)7Bâ86.14 InternVL-U (Tian et al., 2026)1.7B90.3990.7890.6890.2988.7785.18 BAGEL (Deng et al., 2025)7B88.9490.3791.2990.8288.6785.07 Janus-Pro (Chen et al., 2025b)7B86.9088.9089.4089.3289.4884.19 Ovis-U1 (Wang et al., 2025)1.2B82.3790.0888.6893.3585.2083.72 OmniGen2 (Wu et al., 2025b)4B88.8188.8390.1889.3790.2783.57 UniWorld-V1 (Lin et al., 2025a)12B83.6488.3988.4489.2787.2281.38 UniSpace (Ours)8B84.8092.2690.0094.9788.8086.49 supports multimodal understanding, high-quality image editing, and practical text-to-image generation. In particu- lar, it achieves high-quality image editing and competitive text-to-image generation while maintaining a compact 8B model scale, demonstrating the practicality of a unified visual space for large-scale multimodal systems. We release the complete UniSpace system and hope that it will facilitate further research on unified multimodal models. REFERENCES Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025. Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Steven Hoi, Zhaohui Hou, Shijie Huang, Dengyang Jiang, Xin Jin, Liangchen Li, et al. Z-image: An efficient image generation foundation model with single-stream diffusion trans- former. arXiv preprint arXiv:2511.22699, 2025. Mathilde Caron, Hugo Touvron, Ishan Misra, Herv Ě e J Ě egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. Jingjing Chang, Yixiao Fang, Peng Xing, Shuhan Wu, Wei Cheng, Rui Wang, Xianfang Zeng, Gang Yu, and Hai-Bao Chen. Oneig-bench: Omni-dimensional nuanced evaluation for image generation. arXiv preprint arXiv:2506.07977, 2025. Jiuhai Chen, Zhiyang Xu, Xichen Pan, Yushi Hu, Can Qin, Tom Goldstein, Lifu Huang, Tianyi Zhou, Saining Xie, Silvio Savarese, et al. Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset. arXiv preprint arXiv:2505.09568, 2025a. Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Janus-pro: Unified multimodal understanding and generation with data and model scaling.arXiv preprint arXiv:2501.17811, 2025b. 24 UniSpace Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 24185â24198, 2024. Yufeng Cui, Honghao Chen, Haoge Deng, Xu Huang, Xinghang Li, Jirong Liu, Yang Liu, Zhuoyan Luo, Jin- sheng Wang, Wenxuan Wang, et al. Emu3. 5: Native multimodal models are world learners. arXiv preprint arXiv:2510.26583, 2025. Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, Guang Shi, and Haoqi Fan. Emerging properties in unified multimodal pretraining. arXiv preprint arXiv:2505.14683, 2025. Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. In Advances in Neural Information Processing Systems, 2023. Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dock- horn, Jack English, Zion English, Patrick Esser, Sumith Kulal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas M Ě uller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. Flux.1 kontext: Flow matching for in-context image generation and editing in latent space, 2025. URL https://arxiv.org/abs/ 2506.15742. Zongjian Li, Zheyuan Liu, Qihui Zhang, Bin Lin, Feize Wu, Shenghai Yuan, Zhiyuan Yan, Yang Ye, Wangbo Yu, Yuwei Niu, et al. Uniworld-v2: Reinforce image editing with diffusion negative-aware finetuning and mllm implicit feedback. arXiv preprint arXiv:2510.16888, 2025. Bin Lin, Zongjian Li, Xinhua Cheng, Yuwei Niu, Yang Ye, Xianyi He, Shenghai Yuan, Wangbo Yu, Shaodong Wang, Yunyang Ge, et al. Uniworld-v1: High-resolution semantic encoders for unified visual understanding and genera- tion. arXiv preprint arXiv:2506.03147, 2025a. Haokun Lin, Teng Wang, Yixiao Ge, Yuying Ge, Zhichao Lu, Ying Wei, Qingfu Zhang, Zhenan Sun, and Ying Shan.Toklip: Marry visual tokens to clip for multimodal comprehension and generation.arXiv preprint arXiv:2505.05422, 2025b. Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 26296â26306, 2024a. Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-next: Im- proved reasoning, ocr, and world knowledge, January 2024b. URL https://llava-vl.github.io/blog/ 2024-01-30-llava-next/. Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Yingming Wang, Honghao Fu, Chunrui Han, et al. Step1x-edit: A practical framework for general image editing. arXiv preprint arXiv:2504.17761, 2025. Zhiheng Liu, Weiming Ren, Haozhe Liu, Zijian Zhou, Shoufa Chen, Haonan Qiu, Xiaoke Huang, Zhaochong An, Fanny Yang, Aditya Patel, et al. Tuna: Taming unified visual representations for native unified multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 15740â15751, 2026. Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. Chuofan Ma, Yi Jiang, Junfeng Wu, Jihan Yang, Xin Yu, Zehuan Yuan, Bingyue Peng, and Xiaojuan Qi. Unitok: A unified tokenizer for visual generation and understanding. Advances in Neural Information Processing Systems, 38: 129274â129297, 2026. Maxime Oquab, Timoth Ě e Darcet, Th Ě eo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. In Transactions on Machine Learning Research, 2024. 25 UniSpace Liao Qu, Huichao Zhang, Yiheng Liu, Xu Wang, Yi Jiang, Yiming Gao, Hu Ye, Daniel K Du, Zehuan Yuan, and Xinglong Wu. Tokenflow: Unified image tokenizer for multimodal understanding and generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 2545â2555, 2025. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable vi- sual models from natural language supervision. In International Conference on Machine Learning, p. 8748â8763. PMLR, 2021. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj Ě orn Ommer. High-resolution image synthesis with latent diffusion models, 2021. SenseNova-U1 Team. Sensenova-u1: Unifying multimodal understanding and generation with neo-unify architecture. arXiv preprint arXiv:2605.12500, 2026. Jaskirat Singh, Boyang Zheng, Zongze Wu, Richard Zhang, Eli Shechtman, and Saining Xie. Improved baselines with representation autoencoders. arXiv preprint arXiv:2605.18324, 2026. Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. Hao Tang, Chenwei Xie, Xiaoyi Bao, Tingyu Weng, Pandeng Li, Yun Zheng, and Liwei Wang. Unilip: Adapting clip for unified multimodal understanding, generation and editing. arXiv preprint arXiv:2507.23278, 2025. Meituan LongCat Team, Hanghang Ma, Haoxian Tan, Jiale Huang, Junqiang Wu, Jun-Yan He, Lishuai Gao, Songlin Xiao, Xiaoming Wei, Xiaoqi Ma, et al. Longcat-image technical report. arXiv preprint arXiv:2512.07584, 2025. Changyao Tian, Danni Yang, Guanzhou Chen, Erfei Cui, Zhaokai Wang, Yuchen Duan, Penghao Yin, Sitao Chen, Ganlin Yang, Mingxin Liu, et al. Internvl-u: Democratizing unified multimodal models for understanding, reason- ing, generation and editing. arXiv preprint arXiv:2603.09877, 2026. Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. Siglip 2: Multilingual vision-language en- coders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786, 2025. Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, Tianxing Wang, Tianyi Gui, Tingyu Weng, Tong Shen, Wei Lin, Wei Wang, Wei Wang, Wenmeng Zhou, Wente Wang, Wenting Shen, Wenyuan Yu, Xianzhong Shi, Xiaoming Huang, Xin Xu, Yan Kou, Yangyu Lv, Yifei Li, Yijing Liu, Yiming Wang, Yingya Zhang, Yitong Huang, Yong Li, You Wu, Yu Liu, Yulin Pan, Yun Zheng, Yuntao Hong, Yupeng Shi, Yutong Feng, Zeyinzi Jiang, Zhen Han, Zhi-Fan Wu, and Ziyu Liu. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. Guo-Hua Wang, Shanshan Zhao, Xinjie Zhang, Liangfu Cao, Pengxin Zhan, Lunhao Duan, Shiyin Lu, Minghao Fu, Xiaohao Chen, Jianshan Zhao, et al. Ovis-u1 technical report. arXiv preprint arXiv:2506.23044, 2025. Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report. arXiv preprint arXiv:2508.02324, 2025a. Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, et al. Omnigen2: Exploration to advanced multimodal generation. arXiv preprint arXiv:2506.18871, 2025b. Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. In International Conference on Learning Representations, volume 2025, p. 93620â93638, 2025c. 26 UniSpace Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. In International Conference on Learning Representations, volume 2025, p. 28240â28264, 2025. Jinheng Xie, Zhenheng Yang, and Mike Zheng Shou. Show-o2: Improved native unified multimodal models. Advances in Neural Information Processing Systems, 38:47490â47518, 2026. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. Jingfeng Yao, Yuda Song, Yucong Zhou, and Xinggang Wang. Towards scalable pre-training of visual tokenizers for generation. arXiv preprint arXiv:2512.13687, 2025a. Jingfeng Yao, Bin Yang, and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 15703â15712, 2025b. Yang Ye, Xianyi He, Zongjian Li, Bin Lin, Shenghai Yuan, Zhiyuan Yan, Bohan Hou, and Li Yuan. Imgedit: A unified image editing dataset and benchmark. arXiv preprint arXiv:2505.20275, 2025. Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Rep- resentation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. Zhengrong Yue, Haiyu Zhang, Xiangyu Zeng, Boyu Chen, Chenting Wang, Shaobin Zhuang, Lu Dong, Yi Wang, Limin Wang, and Yali Wang. Uniflow: A unified pixel flow tokenizer for visual understanding and generation. arXiv preprint arXiv:2510.10575, 2025. Yue Zhao, Fuzhao Xue, Scott Reed, Linxi Fan, Yuke Zhu, Jan Kautz, Zhiding Yu, Philipp Kr Ě ahenb Ě uhl, and De-An Huang. Qlip: Text-aligned visual tokenization unifies auto-regressive multimodal understanding and generation. arXiv preprint arXiv:2502.05178, 2025. Boyang Zheng, Nanye Ma, Shengbang Tong, and Saining Xie. Diffusion transformers with representation autoen- coders. arXiv preprint arXiv:2510.11690, 2025. 27 UniSpace AABLATION ON BALANCED FLOW MATCHING We ablate the reconstruction-component objective weight Îť r under the same 768+128 unified representation used by our final model. All variants share the same encoder, decoder, training schedule, and sampling setting, isolating the effect of the semanticâreconstruction balance in the flow-matching objective. The sweep shows a clear middle optimum rather than a monotonic preference for larger reconstruction weights. In- creasing the reconstruction objective share from Îť r = 0.25 to Îť r = 0.75 improves FID from 8.99 to 7.10, indicating that under-weighting the reconstruction component leaves decoder-critical visual details insufficiently modeled. How- ever, further increasing the reconstruction share does not continue to improve generation, with Îť r = 0.92 producing 7.39. This supports assigning sufficient but not overwhelming objective weight to the reconstruction component. The best FID is obtained at Îť r = 0.75, matching the weighting used in our final balanced flow-matching objective. Table 12: Ablation of reconstruction-component weighting in balanced flow matching. All variants use the same 768+128 unified representation and are evaluated under the same controlled generation setting. Recon. share Îť r FIDâISâPrecisionâ 0.258.99142.40.681 0.607.42153.20.696 0.757.10154.40.704 0.827.24152.80.705 0.867.13152.50.704 0.927.39150.00.702 BABLATION ON RECONSTRUCTION COMPRESSION We next study the channel compression in the Channel Factorized Merge. This ablation compares our default 128- dimensional reconstruction component with a no-compression variant that preserves the full 768 reconstruction chan- nels. Both variants use the same semantic component and the same factorized concatenation design; only the dimen- sionality of e T r is changed. The no-compression variant confirms the expected reconstruction benefit of a larger reconstruction stream: increasing d r from 128 to 768 improves rFID from 0.163 to 0.085 and PSNR from 29.79 to 33.61. However, this extra detail comes at a clear cost for generation. At both 20 and 40 epochs, the full-dimensional reconstruction component is substantially harder for the DiT prior to model, with FID degrading from 9.95 to 15.51 at epoch 20 and from 6.92 to 11.07 at epoch 40. These results support the role of compression in the Channel Factorized Merge: it is not intended to maximize reconstruction in isolation, but to retain sufficient decoder-critical details while reducing the complexity of the reconstruction distribution that the generative model must learn. Table 13: Ablation of reconstruction-component compression. Reducing the reconstruction stream from 768 to 128 channels slightly weakens pure reconstruction metrics, but substantially improves generative modeling under the same early training budget. Recon. dim d r Unified dim Epoch PSNRâ SSIMâ rFIDâ FIDâ sFIDâISâ Pre.â Rec.â 128768+1282029.790.873 0.1639.957.73 117.5 0.713 0.565 128768+1284029.790.873 0.1636.926.93 144.4 0.730 0.587 768768+7682033.610.940 0.085 15.518.4590.0 0.655 0.574 768768+7684033.610.940 0.085 11.077.80 115.2 0.679 0.596 CENTANGLED REPRESENTATION AND DIAGNOSTIC PROTOCOL We provide additional details of the diagnostic experiment in Sec. 5.1.4. Let I denote an input image, and let Z s = E s (I), Z r = E r (I) 28 UniSpace denote the outputs of a pretrained semantic encoder and a reconstruction-oriented encoder, respectively. We merge the two representations using an MLP: Z m = M([Z s ;Z r ]), where [¡ ;¡] denotes channel-wise concatenation. Unlike the factorized representation used in our method, Z m does not preserve an explicit semanticâreconstruction decomposition. The merger is trained with the same semantic and reconstruction alignment objectives described in the main text. Specifically, Z m is aligned withZ s to preserve semantic information, while a linear projection ofZ m is aligned withZ r to retain reconstruction-relevant information. The semantic decoder D s and the high-fidelity reconstruction decoder D r are trained separately before the diagnostic experiment, with their corresponding encoders kept frozen. At the representation level, the resulting Z m simultaneously supports semantic understanding and high-fidelity reconstruction on real encoded latents. It achieves a zero-shot accuracy of 78.53, compared with 79.10 for the SigLIP baseline, together with a PSNR of 33.83 and an rFID of 0.069. We next train a DiT prior directly on Z m for 20 epochs using the same image distribution and evaluation protocol as in the main experiments. The resulting generated latent exhibits a pronounced decoder-dependent gap: decoding it with the high-fidelity reconstruction decoder D r yields an FID of 120.9, whereas decoding it with the semantic decoder D s yields a much lower FID of 8.07. This gap indicates that the DiT prior primarily learns the semantic-dominant variation in the entangled representation, while failing to model the reconstruction-relevant variation required by D r . Thus, the fact that Z m supports both semantic understanding and high-fidelity reconstruction on real encoded latents does not ensure that both types of information are equally accessible to a generative prior. The reconstruction information is present, but its entangled composition with semantic variation makes it difficult for the generative model to identify and control. To quantify the relative contribution of the semantic pathway, we measure the fraction of the merger output variation explained by the semantic input: Ď s = Var(M(Z s , 0)) Var(M(Z s ,Z r )) . The variance is computed over the evaluation image set using the same merger parameters. We obtain Ď s â 95%, indi- cating that most of the merger output variation is explained by the semantic pathway. This suggests that reconstruction- relevant directions, although sufficient for reconstructing real encoded latents, occupy a relatively small and difficult- to-control part of the entangled latent space. This observation motivates the explicit factorization used in our final representation, where T s and e T r are concatenated along the channel dimension and remain directly accessible during flow-matching training. 29