Fashion Outfit Generation via Unified Sequential Composition Models

# Fashion Outfit Generation via Unified Sequential Composition Models
Source: [https://arxiv.org/html/2608.13888](https://arxiv.org/html/2608.13888)
Kaicheng PANGAffiliation:Laboratory for Artificial Intelligence in DesignAffiliation:Hong Kong SAREmail:[kcpang@aidlab\.hk](mailto:)Xingxing ZOUAffiliation:Hong Kong Polytechnic UniversityAffiliation:Hong Kong SAREmail:[xingxing\.zou@polyu\.edu\.hk](mailto:)Ruohan XUAffiliation:The University of QueenslandAffiliation:Brisbane, AustraliaEmail:[ruohan\.xu@student\.uq\.edu\.au](mailto:)Waikeung WongThanks:Corresponding author\.Affiliation:Laboratory for Artificial Intelligence in DesignAffiliation:Hong Kong SAREmail:[calvinwong@aidlab\.hk](mailto:)

###### Abstract

The task of synthesizing stylistically coherent fashion outfits from massive item libraries, known as fashion outfit generation, remains a non\-trivial challenge, primarily due to the non\-monotonic and implicit nature of aesthetic compatibility, coupled with the exponentially large combinatorial search space\. In this paper, we formalize this task as Constrained Ensemble Generation \(CEG\) and model it as a finite\-horizon deterministic Markov Decision Process\. To address CEG in fashion, we propose the Unified Sequential Composition Model \(USCM\), which jointly models set\-level compatibility and latent composition intents\. Guided by USCM’s learned priors, a Latent Expansion Monte Carlo Tree Search \(LE\-MCTS\) mechanism is proposed to handle item retrieval during composition, balancing local aesthetic synergy with global structural balance\. Extensive experiments on the Polyvore Outfits dataset, along with zero\-shot evaluations on the iFashion and PolyvoreU datasets, demonstrate that our framework achieves state\-of\-the\-art performance across independent human preference evaluations, automated aesthetic proxies, and structural validity metrics for constrained fashion outfit generation\.

## 1Introduction

Fashion outfit generation—the art of assembling garments and accessories into a stylistically coherent ensemble—plays an increasingly important role in modern retail ecosystems\. Yet, generating a high\-quality outfit in practice remains challenging because ensuring stylistic consistency while curating from a large\-scale library leads to a combinatorial explosion of possible solutions\. In this paper, we formalize the general challenge of synthesizing harmonized item collections—such as curating playlists or designing furniture layouts—asConstrained Ensemble Generation \(CEG\)\. Specifically, CEG defines the task of expanding a seed ensemble by selecting items from a massive library to generate a complete ensemble that satisfies both structural requirements and underlying stylistic regularities\. We focus on fashion outfit generation as its representative instantiation, as shown in Figure[1](https://arxiv.org/html/2608.13888#S1.F1)\. Fundamentally, CEG is a Sequential Combinatorial Optimization \(SCO\) problem with exponential search space\. To effectively solve this, we model the generation process as a finite\-horizon, deterministicMarkov Decision Process\(MDP\), where the state is defined as the union of the seed ensemble and all items selected up to the current step\. By representing the state as a collection rather than a sequence, the formulation ensures the process is Markovian while inherently preserving permutation invariance\. Each action corresponds to selecting a new item from the library to be added to the evolving ensemble\. The objective is to navigate this vast search space to identify a terminal state that maximizes a holistic, constraint\-aware compatibility function\. However, addressing this specific MDP presents three distinct challenges: \(1\)Combinatorial Explosion\. Real\-world fashion libraries function as large\-scale discrete repositories, leading to a combinatorial explosion of possible ensembles that is computationally intractable for exhaustive search or standard combinatorial optimization methods\. \(2\)Non\-monotonic Latent Objective\. The fashion compatibility function lacks a closed\-form expression and must be learned from data\. It is inherently non\-monotonic, as adding a single item may improve the ensemble locally while compromising its global harmony\. \(3\)Sparse and Delayed Reward Signal\. In mainstream fashion datasets, human\-curated ensembles constitute only a tiny fraction of the vast combinatorial space\. Since ground\-truth labels only exist for these rare completed outfits, partial states lack direct supervision to evaluate their intermediate aesthetic compatibility during ensemble construction\.

![Refer to caption](https://arxiv.org/html/2608.13888v1/comparsion_between_traditional_methods_and_my_methods.png)Figure 1:Illustration of Constrained Ensemble Generation \(CEG\) in fashion context\. Given a seed ensembleS0S\_\{0\}and a large\-scale library𝒳\\mathcal\{X\}, the goal is to synthesize a coherent outfit that maximizes constraint\-aware compatibility objective\. TraditionalDiscriminative Scoringapproaches \(red path\) lack predictive policy guidance for next\-item intent, forcing brute\-force retrieval across candidate libraries that often results in structural or stylistic failure\. In contrast, our proposedUSCM\(green path\) leverages learned policy priors and value proxies to generate coherent outfits\.Previous research in neural fashion recommendation has largely sidestepped these core challenges\. Early studies mainly focus on assessing the aesthetic compatibility of pre\-composed outfits\([10](https://arxiv.org/html/2608.13888#bib.bib10);[21](https://arxiv.org/html/2608.13888#bib.bib9);[18](https://arxiv.org/html/2608.13888#bib.bib12);[8](https://arxiv.org/html/2608.13888#bib.bib11)\), supporting selection among static candidates rather than generating new ensembles\. More recent approaches[14](https://arxiv.org/html/2608.13888#bib.bib8);[20](https://arxiv.org/html/2608.13888#bib.bib7);[1](https://arxiv.org/html/2608.13888#bib.bib17)move toward complementary item recommendation, where the goal is to identify a single item to complete a nearly finished outfit\. While these paradigms provide useful foundations, they remain limited in practical outfit generation: the former is essentially retrospective filtering, and the latter reduces outfit generation to a greedy completion process\. Neither foresees the long\-term aesthetic consequences of sequential actions in a sparse\-reward environment\.

To bridge this gap, we introduce a representation\-driven ensemble synthesis framework built on theUnified Sequential Composition Model \(USCM\)\. Specifically, USCM jointly learns compatibility evaluation and policy priors, paired with Latent Expansion Monte Carlo Tree Search \(LE\-MCTS\) to perform forward\-looking candidate selection via dynamic latent branch expansion for the CEG problem\. Serving as the robust representation engine of our framework, USCM translates the current ensemble state into actionable search guidance via a dual\-pathway Transformer architecture\. Specifically, aValue Headestimates an ensemble’s compatibility score \(v^t\\hat\{v\}\_\{t\}\) to provide dense aesthetic feedback, while aPolicy Headprojects a query ensemble into a continuous latent intent \(𝐳t\\mathbf\{z\}\_\{t\}\) to guide candidate discovery within the latent manifold\. Additionally, USCM enables embedding library items into the shared manifold to facilitate efficient retrieval\. Unlike previous pipelines\([20](https://arxiv.org/html/2608.13888#bib.bib7)\)that rely on separate models for these capabilities, USCM employs a multi\-task learning paradigm\. This joint optimization leverages task synergy to capture deep stylistic correlations while reducing parameter redundancy\. To complement USCM during generation, we introduce Latent Expansion Monte Carlo Tree Search \(LE\-MCTS\) to handle set synthesis over massive item libraries\. Instead of enumerating discrete library items, LE\-MCTS uses USCM’s policy intent𝐳t\\mathbf\{z\}\_\{t\}for top\-kklatent retrieval and its value scorev^t\\hat\{v\}\_\{t\}for branch assessment\. Given a simulation budgetMM, LE\-MCTS evaluates potential item combinations to avoid short\-sighted selections, ensuring that the synthesized ensemble satisfies both aesthetic and structural constraints\.

Extensive experiments on three mainstream datasets\-Polyvore Outfits, iFashion, and PolyvoreU\-empirically validate our approach\. Critically, USCM establishes new state\-of\-the\-art results on fundamental discriminative tasks demonstrating superior representational capacity as a core heuristic engine\. In terms of fashion generation task, our framework consistently surpasses greedy baselines and retrieval\-based methods by a clear margin\. Furthermore, evaluations on structural metrics confirm that our approach best satisfies structural rules without relying on explicit structural rules\. Comprehensive ablations highlight the advantages of the unified architecture, search strategies and budgets, and the framework’s robust zero\-shot transferability to unseen data\. Our main contributions are summarized as follows:

- •We formalize Constrained Ensemble Generation \(CEG\) as a deterministic MDP, providing an unified problem formulation to address combinatorial explosion, non\-monotonic objective, and sparse feedback in set synthesis, with fashion outfit generation as an instantiation\.
- •We propose the Unified Sequential Composition Model \(USCM\), a multi\-task representation network that unifies outfit compatibility evaluation and next\-item latent intent prediction within a shared multimodal manifold for ensembles and items\. During generation, it is coupled with a LE\-MCTS mechanism for lookahead candidate selection\.
- •Extensive evaluations demonstrate that our framework achieves state\-of\-the\-art performance across both fundamental discriminative benchmarks and the CEG task, validating its ability to navigate complex aesthetic landscapes even in zero\-shot scenarios\.

## 2Related Work

Earlier studies addressing fashion outfit generation have focused on discriminative tasks, using metric learning\([10](https://arxiv.org/html/2608.13888#bib.bib10)\), sub\-space projections\([22](https://arxiv.org/html/2608.13888#bib.bib5)\), and graph neural network\([9](https://arxiv.org/html/2608.13888#bib.bib20)\)to predict outfit compatibility\. The field has further evolved to incorporate large multimodal models\([5](https://arxiv.org/html/2608.13888#bib.bib15)\)and graph attention networks\([19](https://arxiv.org/html/2608.13888#bib.bib16)\)\. However, these methods remain passive evaluators and cannot navigate an expensive search space for ensemble composition\. Recently, generative methods\([21](https://arxiv.org/html/2608.13888#bib.bib9);[20](https://arxiv.org/html/2608.13888#bib.bib7)\)have emerged to address complementary item retrieval task, typically by predicting latent embeddings to retrieve items from a library\. Some approaches also utilize GNNs for selection logic\([3](https://arxiv.org/html/2608.13888#bib.bib13)\), and image\-to\-image translation to generate retrieval templates\([1](https://arxiv.org/html/2608.13888#bib.bib17)\)\. However, these methods are primarily designed for one\-step recommendation and lack the foresight of each selection\. Thus, their reliance on greedy decisions often leads to stylistic inconsistency in terms of generating a sequence of items\. Another branch of generative research[25](https://arxiv.org/html/2608.13888#bib.bib21);[23](https://arxiv.org/html/2608.13888#bib.bib14);[24](https://arxiv.org/html/2608.13888#bib.bib18)focuses on synthesizing a set of real\-photo items directly\. These methods rely on input constraints such as reference masks or explicit category prompt\. In contrast, CEG requires the model to navigate a massive unstructured library without any auxiliary structural guidance\. Thus, we exclude them from our comparative analysis\. A parallel line of research explores sequential decision\-making within the broader fashion domain, though predominantly for operational modeling rather than generative tasks\. For instance, MDPs have been deployed to optimize dynamic pricing for seasonal products\([2](https://arxiv.org/html/2608.13888#bib.bib24)\)and to manage attribute\-level inventory risk across color variants\([11](https://arxiv.org/html/2608.13888#bib.bib22)\)\.\([12](https://arxiv.org/html/2608.13888#bib.bib23)\)integrated cognitive frameworks and decision\-making models to predict human fashion choices based on contextual choice sets\. These prior works operate over structured, low\-dimensional state and action spaces, where decisions correspond to finite operational adjustments\. In contrast, CEG necessitates search\-based planning to navigate a combinatorially explosive search space, thereby sequentially optimizing a non\-monotonic latent aesthetic landscape\.

## 3Methodology

### 3\.1Problem Formulation

We formulateConstrained Ensemble Generation\(CEG\) as a finite\-horizon deterministicMarkov Decision Process\. Let𝒳\\mathcal\{X\}denote a large\-scale library of fashion items\. To preserve permutation invariance, the state space𝒮\\mathcal\{S\}is defined over the power set of the library,i\.e\.,𝒮⊆𝒫⁡\(𝒳\)∖\{∅\}\\mathcal\{S\}\\subseteq\\mathcal\{P\}\(\\mathcal\{X\}\)\\setminus\\\{\\emptyset\\\}\. An episode starts from a seed ensembleS0⊂𝒳S\_\{0\}\\subset\\mathcal\{X\}with\|S0\|≥1\|S\_\{0\}\|\\geq 1\. At each discrete steptt, the agent selects an itemAt∈𝒳∖StA\_\{t\}\\in\\mathcal\{X\}\\setminus S\_\{t\}, inducing the deterministic transition:St\+1=St∪\{At\}S\_\{t\+1\}=S\_\{t\}\\cup\\\{A\_\{t\}\\\}\. AfterTTsteps, the process reaches a terminal stateSTS\_\{T\}\. A key aspect of CEG is that the notion ofconstraintsconsists of two complementary components\. The first is a set ofexplicit constraints, denoted by𝒞exp\\mathcal\{C\}^\{\\mathrm\{exp\}\}, comprising the seed ensemble, item uniqueness, and cardinality limits\. These constraints define the feasible solution space\. In contrast, the second is a set oflatent constraints, denoted by𝒞lat\\mathcal\{C\}^\{\\mathrm\{lat\}\}, which capture the stylistic and structural regularities underlying fashion ensembles\. Unlike explicit constraints, these regularities are not available in closed form and cannot be exhaustively encoded by hand\-crafted rules; rather, they must be inferred from data\. Under this formulation, intermediate state transitions yield zero immediate rewards, with the overall compatibility score evaluated solely at the terminal state\.

Accordingly, CEG aims to identify the optimal terminal ensembleST∗S\_\{T\}^\{\*\}that maximizes its compatibility score within the feasible solution space:

ST∗=arg⁡maxST∈Ω⁡\(S0,𝒞exp\)⁡Φ⁡\(ST,𝒞lat\),S\_\{T\}^\{\*\}=\\arg\\max\_\{S\_\{T\}\\in\\Omega\(S\_\{0\};\\mathcal\{C\}^\{\\mathrm\{exp\}\}\)\}\\Phi\(S\_\{T\};\\mathcal\{C\}^\{\\mathrm\{lat\}\}\),\(1\)whereΦ⁡\(⋅,𝒞lat\):𝒮→ℝ\\Phi\(\\cdot;\\mathcal\{C\}^\{\\mathrm\{lat\}\}\):\\mathcal\{S\}\\rightarrow\\mathbb\{R\}is an implicit compatibility scoring mapping induced from latent constraints, andΩ⁡\(S0,𝒞exp\)\\Omega\(S\_\{0\};\\mathcal\{C\}^\{\\mathrm\{exp\}\}\)denotes the set of all admissible ensembles reachable fromS0S\_\{0\}under the explicit constraints\. The core complexity of CEG stems from its dual constraint structure: while explicit constraints merely delineate the admissible state space, the critical stylistic regularities remain implicit, non\-monotonic, and lack a closed\-form analytical specification\. Consequently, identifying the optimal ensembleST∗S\_\{T\}^\{\*\}within this paradigm is computationally intractable for exhaustive search, given the combinatorially explosive nature of the search space\. To solve this complex problem, we train an unified model to supply state value estimations and policy priors, which coupled with a tree search mechanism to compose the near\-optimal stateS^T\\hat\{S\}\_\{T\}\.

![Refer to caption](https://arxiv.org/html/2608.13888v1/model_structure.png)Figure 2:Architecture of the Unified Sequential Composition Model\. \(a\) The Value Head estimates the latent potential of an ensemble to guide long\-term composition, while the Policy Head projects the ensemble into a latent intent𝐳t\\mathbf\{z\}\_\{t\}to induce a policy prior over library\. \(b\) Library items are indexed into the same manifold via shared weights\. This shared\-weight architecture ensures efficient heuristic\-guided retrieval and stylistic consistency between two heads\.
### 3\.2Unified Sequential Composition Model

To address CEG, we propose theUnified Sequential Composition Model \(USCM\)as the foundational multi\-task neural architecture of our framework\. As discussed in Section[3\.1](https://arxiv.org/html/2608.13888#S3.SS1), solving CEG primarily requires modeling the latent stylistic constraints encoded by the implicit compatibility objectiveΦ\\Phi\. To this end, USCM is designed to learn a shared representational manifold where evolving ensembles, next\-item intents, and individual item embeddings are jointly aligned\. As illustrated in Figure[2](https://arxiv.org/html/2608.13888#S3.F2), the USCM processes the current stateStS\_\{t\}by prepending a learnable\[Task Token\]to the frozen CLIP embeddings of the constituent items\. The Transformer encoder aggregates them into a global sequence representationeseqe\_\{\\mathrm\{seq\}\}, which is then branched into two specialized components:

Value Head\.The functionfv:𝒮→ℝf\_\{v\}:\\mathcal\{S\}\\rightarrow\\mathbb\{R\}serves as aoutfit compatibility evaluator\. Specifically, it is trained exclusively to predict the overall aesthetic score of complete ensembles,Φ⁡\(ST,𝒞lat\)\\Phi\(S\_\{T\};\\mathcal\{C\}^\{\\mathrm\{lat\}\}\)\. Crucially, because USCM’s Transformer architecture is length\-agnostic and permutation\-invariant, this terminal evaluator can approximately process intermediate statesStS\_\{t\}\(t<Tt<T\)\. Consequently,fvf\_\{v\}acts as a zero\-shot heuristic proxy during inference, directly outputting an aesthetic compatibility score for partial ensembles\. This mechanism provides the LE\-MCTS decoding process with dense, immediate feedback for candidate assessment—bypassing computationally expensive simulation rollouts and effectively resolving the fundamental bottleneck of sparse and delayed rewards\. Policy Head\.To address the combinatorial explosion of the action space𝒳\\mathcal\{X\}, the Policy Headfp:𝒮→ℝdf\_\{p\}:\\mathcal\{S\}\\rightarrow\\mathbb\{R\}^\{d\}maps the current stateStS\_\{t\}into a latent query𝐳t\\mathbf\{z\}\_\{t\}representing the stylistic intent—a continuous\-space prior that encodes the representational target for the next selection\. By embedding both states and library items into a unified manifold,fpf\_\{p\}induces a policy prior via nearest\-neighbor retrieval, thereby enabling selective expansion by effectively pruning the massive search space\.

Crucially, every itemi∈𝒳i\\in\\mathcal\{X\}is projected into the same latent manifold via the shared encoder and linear layer to obtain an indexing embedding𝐳i∈ℝd\\mathbf\{z\}\_\{i\}\\in\\mathbb\{R\}^\{d\}\. This unified design ensures that the predicted intent𝐳t\\mathbf\{z\}\_\{t\}and candidate items from the library are semantically aligned\. By consolidating compatibility estimation and policy prior generation into a single architecture, USCM reduces parameter redundancy and benefits from multi\-task learning, as empirically validated in Section[4\.4](https://arxiv.org/html/2608.13888#S4.SS4)\.

### 3\.3Training Objectives

We jointly optimize USCM for two complementary tasks:compatibility value estimationandpolicy prior learning\. The total training objective isℒ=ℒv\+λp​ℒp\\mathcal\{L\}=\\mathcal\{L\}\_\{v\}\+\\lambda\_\{p\}\\mathcal\{L\}\_\{p\}\.

#### Compatibility Value Estimation \(ℒv\\mathcal\{L\}\_\{v\}\) via Robust Negative Sampling\.

This objective trains USCM to approximate the terminal compatibility functionalΦ⁡\(⋅,𝒞lat\)\\Phi\(\\cdot;\\mathcal\{C\}^\{\\mathrm\{lat\}\}\)induced by latent stylistic constraints\. We formulate this task as a binary classification task that distinguishes positive ensembles from hard negatives, and optimize it using Focal Loss[13](https://arxiv.org/html/2608.13888#bib.bib1)\. Specifically, we implementrobust negative samplingstrategy by adopting a1:2:21\{:\}2\{:\}2positive\-to\-negative ratio, where negative samples are constructed in two ways: \(i\) replacing items with alternatives from the same category, and \(ii\) grouping visually similar items drawn from the library into a single ensemble\. These hard negatives expose the model to subtle violations of latent stylistic consistency, forcing the Value Head to capture the non\-monotonic compatibility landscape underlyingΦ⁡\(⋅,𝒞lat\)\\Phi\(\\cdot;\\mathcal\{C\}^\{\\mathrm\{lat\}\}\)rather than relying on visual similarity or category co\-occurrence alone\.

#### Latent Intent Alignment \(ℒp\\mathcal\{L\}\_\{p\}\) via Contrastive Learning\.

To guide the search through the massive action space𝒳\\mathcal\{X\}, the Policy Head learns to represent the latent stylistic intent\. Given a partial ensembleStS\_\{t\}, the model produces a latent query𝐳t=fp​\(St\)\\mathbf\{z\}\_\{t\}=f\_\{p\}\(S\_\{t\}\), which is trained to align with the embedding of a positive counterpart—an item sampled from the same human\-curated outfit—denoted as𝐳t\+1\\mathbf\{z\}\_\{t\+1\}\. We optimize this alignment using the InfoNCE loss[17](https://arxiv.org/html/2608.13888#bib.bib2)\. For a batch ofBBtraining pairs, the objective is:ℒp=−1B∑k=1Blogexp⁡\(sim​\(𝐳t,k,𝐳t\+1,k\)/τ\)∑j=1Bexp⁡\(sim​\(𝐳t,k,𝐳t\+1,j\)/τ\)\\mathcal\{L\}\_\{p\}=\-\\frac\{1\}\{B\}\\sum\_\{k=1\}^\{B\}\\log\\frac\{\\exp\(\\text\{sim\}\(\\mathbf\{z\}\_\{t,k\},\\mathbf\{z\}\_\{t\+1,k\}\)/\\tau\)\}\{\\sum\_\{j=1\}^\{B\}\\exp\(\\text\{sim\}\(\\mathbf\{z\}\_\{t,k\},\\mathbf\{z\}\_\{t\+1,j\}\)/\\tau\)\}, wheresim​\(⋅,⋅\)\\text\{sim\}\(\\cdot,\\cdot\)denotes cosine similarity andτ\\tauis a temperature parameter\. Unlike the triplet loss, InfoNCE provides a stronger discriminative signal by treating all other items in the batch as negatives\. Jointly optimizingℒv\\mathcal\{L\}\_\{v\}andℒp\\mathcal\{L\}\_\{p\}yields a unified representation in which state evaluation and next\-item prediction mutually reinforce each other\. By mapping both partial ensembles and library items into a unified metric manifold, USCM enables the LE\-MCTS to efficiently retrieve high\-prior candidates via nearest\-neighbor search, effectively pruning the search tree from the outset\.

### 3\.4Latent Expansion Monte Carlo Tree Search

We introduce theLatent Expansion Monte Carlo Tree Search \(LE\-MCTS\), a planning algorithm built on top of the learned heuristics provided by USCM\. The detailed procedure for LE\-MCTS is summarized in Appendix[A](https://arxiv.org/html/2608.13888#A1)\. Under the explicit constraints𝒞exp\\mathcal\{C\}^\{\\mathrm\{exp\}\}, the search process is restricted to feasible states, while the latent constraints𝒞lat\\mathcal\{C\}^\{\\mathrm\{lat\}\}are handled indirectly through the value and policy estimates produced by USCM\. In this sense, our framework does not assume that the full constraint structure is analytically available; instead, it performs look\-ahead search in the explicit feasible space while using learned latent compatibility signals to evaluate and expand promising branches\. Unlike conventional MCTS, which assumes a fixed and explicitly enumerable discrete action space, our setting involves a massive item library𝒳\\mathcal\{X\}for which exhaustive expansion is impractical\. LE\-MCTS addresses this issue by dynamically constructing the local action set from the latent manifold learned by USCM\. Each nodeuucorresponds to a stateSuS\_\{u\}and stores visit statistics\(nu,wu,qu\)\(n\_\{u\},w\_\{u\},q\_\{u\}\), together with a prioritized candidate poolPuP\_\{u\}containing retrieved items and their associated policy priors\.

#### Selection and Progressive Widening\.

From the root, LE\-MCTS traverses the tree by recursively selecting the childccthat maximizes the PUCT objective\([4](https://arxiv.org/html/2608.13888#bib.bib4)\):PUCT⁡\(c\)=qc\+Cpuct⋅pc​nparent1\+nc\\mathrm\{PUCT\}\(c\)=q\_\{c\}\+C\_\{\\mathrm\{puct\}\}\\cdot p\_\{c\}\\frac\{\\sqrt\{n\_\{\\mathrm\{parent\}\}\}\}\{1\+n\_\{c\}\}, wherepc=πθ​\(a∣Su\)p\_\{c\}=\\pi\_\{\\theta\}\(a\\mid S\_\{u\}\)denotes the policy prior assigned to actionaaat stateSuS\_\{u\},θ\\thetadenotes the learned parameters of the USCM, andCpuctC\_\{\\mathrm\{puct\}\}controls the exploration–exploitation trade\-off\. To avoid excessive branching in the large action space, we adopt Progressive Widening\([15](https://arxiv.org/html/2608.13888#bib.bib3)\)\. A new child is expanded only when\|children⁡\(u\)\|<Cpw⋅\(nu\)αpw\|\\mathrm\{children\}\(u\)\|<C\_\{\\mathrm\{pw\}\}\\cdot\(n\_\{u\}\)^\{\\alpha\_\{\\mathrm\{pw\}\}\}\.CpwC\_\{\\mathrm\{pw\}\}andαpw\\alpha\_\{\\mathrm\{pw\}\}regulate the expansion rate\.

Latent Expansion and Value Estimation\.When a nodeuuis first visited or selected for widening, the USCM’s policy head generates a latent query𝐳u=fp​\(Su\)\\mathbf\{z\}\_\{u\}=f\_\{p\}\(S\_\{u\}\)\. We then retrieve a candidate set𝒜u\\mathcal\{A\}\_\{u\}of the top\-kkitems from the library𝒳\\mathcal\{X\}based on their latent similarity to𝐳u\\mathbf\{z\}\_\{u\}:𝒜u=\{Ai∣𝐳i∈Top\-​k​\(𝒟,𝐳u\)\}\\mathcal\{A\}\_\{u\}=\\\{A\_\{i\}\\mid\\mathbf\{z\}\_\{i\}\\in\\text\{Top\-\}k\(\\mathcal\{D\},\\mathbf\{z\}\_\{u\}\)\\\}, where𝒟=\{zi\}i∈𝒳\\mathcal\{D\}=\\\{z\_\{i\}\\\}\_\{i\\in\\mathcal\{X\}\}is the pre\-indexed latent library built by the USCM item\-indexing pathway\. The priorp⁡\(Ai\|Su\)p\(A\_\{i\}\|S\_\{u\}\)is derived by applying a tempered Softmax to the cosine similarities between𝐳u\\mathbf\{z\}\_\{u\}and𝐳i\\mathbf\{z\}\_\{i\}\. The highest\-priority feasible action is then selected for expansion,Au∗=arg⁡maxAi∈Au⁡p⁡\(Ai∣Su\)A\_\{u\}^\{\*\}=\\arg\\max\_\{A\_\{i\}\\in A\_\{u\}\}p\(A\_\{i\}\\mid S\_\{u\}\), and the corresponding child state is formed asSc=Su∪\{Au∗\}S\_\{c\}=S\_\{u\}\\cup\\\{A\_\{u\}^\{\*\}\\\}\. The Value Head evaluates the newly expanded child asv^c=fv​\(Sc\)\\hat\{v\}\_\{c\}=f\_\{v\}\(S\_\{c\}\), which serves as a leaf evaluation estimating the compatibility potential of this intermediate state under the latent stylistic constraints𝒞lat\\mathcal\{C\}^\{\\mathrm\{lat\}\}\.

#### Backpropagation\.

The evaluated valuev^c\\hat\{v\}\_\{c\}is then propagated back along the traversal path to the root\. For each nodeuuon this path, we updatenu←nu\+1n\_\{u\}\\leftarrow n\_\{u\}\+1,wu←wu\+v^cw\_\{u\}\\leftarrow w\_\{u\}\+\\hat\{v\}\_\{c\}, andqu←wu/nuq\_\{u\}\\leftarrow w\_\{u\}/n\_\{u\}\. AfterMMsimulations from stateStS\_\{t\}, the agent executes the action corresponding to the child of the root with the highest visit count\([4](https://arxiv.org/html/2608.13888#bib.bib4)\)\. By combining latent retrieval with look\-ahead planning, our method converts the globally difficult CEG objective into a sequence of locally tractable search decisions\.

### 3\.5End\-to\-End Generation Pipeline

To synthesize a complete ensemble during inference, we integrate the aforementioned components into an iterative, search\-based pipeline\. The process initializes with a seed ensembleS0S\_\{0\}\. At each decision steptt, we executes a budget ofMMLE\-MCTS simulations\. During this phase, the USCM serves as the heuristic engine, mapping intermediate states to latent intents to dynamically retrieve candidates, and providing value estimations to guide branch expansion\. Once the simulations conclude, we select the next itemAtA\_\{t\}—corresponding to the child node with the highest visit count—and then transit to the new stateSt\+1=St∪\{At\}S\_\{t\+1\}=S\_\{t\}\\cup\\\{A\_\{t\}\\\}\. To determine the terminal state without rigid templates, we introduce aMarginal Gain\-based Terminal Operatorguided by cardinality limits in𝒞exp\\mathcal\{C\}^\{\\mathrm\{exp\}\}\. Specifically, the generation process halts when the incremental aesthetic gain,Δ​v^=v^​\(St\+1\)−v^​\(St\)\\Delta\\hat\{v\}=\\hat\{v\}\(S\_\{t\+1\}\)\-\\hat\{v\}\(S\_\{t\}\), drops below a predefined tolerance thresholdϵ\>0\\epsilon\>0, signaling aesthetic saturation\. This allows dynamic ensemble sizing governed by stylistic necessity rather than hand\-crafted rules\.

## 4Experiments

### 4\.1Experimental Setup

We evaluate our model on the non\-disjoint version of the Polyvore Outfits dataset\([22](https://arxiv.org/html/2608.13888#bib.bib5)\)\(53,306 train/10,000 test outfits\), encompassing 251,008 fashion items\. We also conduct a zero\-shot ablation study on the iFashion\([6](https://arxiv.org/html/2608.13888#bib.bib19)\)and PolyvoreU dataset\([16](https://arxiv.org/html/2608.13888#bib.bib25)\), which comprises 49,357 and 178,481 outfits, respectively\. The implementation details are provided in Appendix[B](https://arxiv.org/html/2608.13888#A2)\.

### 4\.2Main Results: Constrained Ensemble Generation Quality

Baseline Configurations\.We first compare our framework against four representative baselines on the CEG task: \(1\)Random, which arbitrarily selects candidates from the library to establish an empirical performance lower bound; \(2\)Type\-aware\([22](https://arxiv.org/html/2608.13888#bib.bib5)\), which lacks next\-item intent prediction and instead greedily ranks200200randomly sampled candidates at each step via its discriminative compatibility scorer; \(3\)VLLMleverages Gemini\-3\.0\-flash to generate descriptive queries for complementary items using images of the seed ensemble as visual prompts; and \(4\)OutfitTransformer\([20](https://arxiv.org/html/2608.13888#bib.bib7)\)utilizes its next\-item prediction capability to select item at each step\.

Metrics\.We evaluate the generated ensembles across two dimensions:Aesthetic CompatibilityandStructural metrics\.Since the true aesthetic compatibility functionΦ⁡\(ST,Clat\)\\Phi\(S\_\{T\};C^\{\\text\{lat\}\}\)is a subjective and unobservable black\-box, any single neural evaluator inevitably introduces its own inductive bias\. To alleviate the evaluation bias, we conduct a blind A/B human preference study along with a neural evaluator, treating them as independent evaluators to approximateΦ\\Phi: \(1\)Human Pref\.reports win rates from an A/B blind human study, where3232independent evaluators assessed3030randomized test cases, yielding a total of928928paired choices\. \(2\)SneuralS\_\{\\text\{neural\}\}is an independent neural compatibility scorer adopting the Transformer\-based architecture from OutfitTransformer\([20](https://arxiv.org/html/2608.13888#bib.bib7)\), which outputs unnormalized raw logits in\(−∞,\+∞\)\(\-\\infty,\+\\infty\)to quantify holistic ensemble harmony\. Furthermore, to measure how well the model recovers the underlyingfashion grammarwithout explicit templates, we employ two structural metrics detailed in Appendix[C](https://arxiv.org/html/2608.13888#A3): \(1\)Category Jensen\-Shannon Divergence \(Cate\. JSD\)measures the divergence between the category distribution of generated outfits and the empirical ground\-truth distribution\. A lower JSD indicates superior alignment with real\-world category proportions\. This prevents the evaluation from being biased toward models that exploit the compatibility metric through category inflation, adding items without regard for stylistic coherence\. \(2\)SvalidS\_\{\\text\{valid\}\}is a rule\-based metric that audits the structural validity of generated outfits\. An ensemble is considered valid only if it satisfies both functional coverage and logical compositional constraints\.

Table 1:Performance evaluation of the CEG task on Polyvore Outfits dataset\. Our proposed framework consistently achieves state\-of\-the\-art performance in both aesthetic compatibility and structural metrics\. Reported values for stochastic methods are means over three independent runs\.Analysis of Constrained Ensemble Generation Results\.As summarized in Table[1](https://arxiv.org/html/2608.13888#S4.T1), our framework achieves state\-of\-the\-art performance on the CEG task across aesthetic compatibility and structural metrics exceptSvalidS\_\{\\text\{valid\}\}\. Regarding aesthetic quality, human evaluators significantly preferred our generated outfits \(41\.81%\) over VLLM \(32\.65%\) and OutfitTransformer \(25\.54%\) \(χ2,p<0\.01\\chi^\{2\},p<0\.01\)\. On the neural aesthetic metric, our method similarly achieves state\-of\-the\-art performance with anSneuralS\_\{\\text\{neural\}\}of13\.4313\.43, outperforming OutfitTransformer \(11\.6211\.62\) and VLLM \(5\.285\.28\)\. In terms of structural metrics, our method achieves the lowest Category JSD \(0\.00590\.0059\)\. Since Cate\. JSD measures the divergence between the generated category distribution and the ground\-truth dataset distribution, a lower score indicates higher fidelity to natural clothing composition\. Notably, our model outperforms even theRandombaseline \(0\.00700\.0070\)—which inherently mimics dataset frequencies—proving that USCM captures the implicit fashion grammar without collapsing into shortcut strategies \(e\.g\., over\-generating safe categories like accessories\)\.

Among the comparative baselines, theRandommethod yields the lowest performance, confirming the inherent complexity and non\-triviality of the CEG task\. WhileType\-awareimproves local compatibility through a discriminative scorer, its step\-wise greedy ranking from blind samples restricts its ability to achieve long\-horizon structural coordination\. Conversely,VLLMachieves the highest functional completeness \(Svalid=0\.845S\_\{\\text\{valid\}\}=0\.845\), owed to the extensive visual\-language commonsense in LLMs\. However, it suffers from severe category distribution shift \(JSD=0\.0198\\text\{JSD\}=0\.0198\) and degraded aesthetic quality, highlighting a semantic\-to\-physical retrieval mismatch and a lack of fine\-grained visual compatibility alignment inherent in current VLLM agents\.OutfitTransformerexhibits a similar structural degradation \(JSD=0\.0183\\text\{JSD\}=0\.0183\), indicating that its next\-item completion strategy sacrifices global set balance\. Qualitative comparisons and generating variations are provided in Appendix[D\.1](https://arxiv.org/html/2608.13888#A4.SS1)and[D\.2](https://arxiv.org/html/2608.13888#A4.SS2), respectively\.

### 4\.3Evaluation on Downstream Discriminative Tasks

To verify the representation capability of USCM as the core heuristic engine for planning, we evaluate its performance across three standard discriminative tasks:Compatibility Prediction \(Comp\.\), which evaluates the model’s ability to discriminate between ground\-truth and negative ensembles using the Area Under the Curve \(AUC\) metric;Fill\-In\-The\-Blank \(FITB\), which tests model’s capability to select the most compatible item from four choices to complete a partial outfit; andComplementary Item Retrieval \(R@K\), which measures the ranking of ground\-truth items against 3,000 distractors given a query outfit\. Table[2](https://arxiv.org/html/2608.13888#S4.T2)summarizes the results on two versions of the Polyvore Outfits dataset\. USCM consistently achieves strong performance, outperforming all previous SOTA methods across every metric\. On the disjoint set, USCM achieves a compatibility AUC of0\.960\.96, a significant absolute improvement of0\.080\.08over OutfitTransformer \(0\.880\.88\)\. Similarly, in the FITB task, our model reaches70\.41%70\.41\\%accuracy, surpassing the previous best results by over10%10\\%\. In the more challenging retrieval task, USCM nearly doubles the R@10 performance compared to CSA\-Net \(11\.77%11\.77\\%vs\.5\.93%5\.93\\%\), demonstrating the superior alignment of our latent space for large\-scale complementary item retrieval\. Furthermore, we investigate the impact of our Robust Negative Sampling \(RNS\) strategy, described in Section[3\.3](https://arxiv.org/html/2608.13888#S3.SS3.SSS0.Px1)\. ComparingUSCMwithUSCM\-w/o\-RNS, we observe that incorporating RNS strategy provides a consistent gains across both splits\. Specifically, RNS improves disjoint FITB accuracy from68\.49%68\.49\\%to70\.41%70\.41\\%and substantially bolsters R@50, indicating that hard negatives effectively sharpen the model’s decision boundaries\.

Table 2:Evaluation on downstream discriminative tasks on the Polyvore Outfits dataset\. USCM consistently outperforms baselines across all metrics\. USCM\-w/o\-RNS denotes a variant of USCM trained without the robust negative sampling strategy\.
### 4\.4Ablation Study

Ablation on Model Architectures\.Comparing the unified USCM against a separate baseline, where the policy and value heads are trained as independent models, confirms that joint multi\-task learning yields more robust stylistic representations\. As shown in Table[3](https://arxiv.org/html/2608.13888#S4.T3), the unified architecture demonstrates superior performance across almost all metrics, confirming that joint multi\-task learning yields more robust stylistic representations\. In downstream discriminative tasks, the USCM significantly outperforms the separate model in FITB and Retrieval metrics \(3\.45%3\.45\\%gain in FITB and2\.95%2\.95\\%in R@10\)\. Notably, while both models achieve comparable scores on theComp\.metric, the substantial lead in FITB task indicates that the USCM provides far more accurate guidance for the LE\-MCTS algorithm to identify promising candidates during the search\. This advantage translates into a substantial improvement in structural integrity\. The USCM achieves a significant lead in both category JSD \(0\.00590\.0059vs\.0\.01340\.0134\) and ensemble validity \(0\.8150\.815vs\.0\.5090\.509\)\. The diminished validity score of the separate model suggests that decoupling scoring and selection capabilities may induce conflicting latent spaces, thereby potentially hindering the search process from converging on logically sound ensembles\. USCM’s stable performance across diverse, independent evaluators proves that the unified architecture is a more robust heuristic engine for CEG\.

Table 3:Performance comparison betweenunifiedandseparatesequential composition model\.Ablation on Search Strategies\.To isolate the specific contribution of LE\-MCTS from USCM’s underlying representation capacity, we compare LE\-MCTS against alternative search strategies—Beam Search \(Beam width is 10\) and Reranking \(Search width is 10\)—under identical 10\-size candidate retrieval pools, USCM priors, and computational budgets\. To ensure objective evaluation, five domain experts assessed 10 generated ensembles across 5\-point Likert scales for Aesthetic Compatibility and Structural Completeness alongside a blind preference vote\. As detailed in Table[4](https://arxiv.org/html/2608.13888#S4.T4), LE\-MCTS achieves superior performance across all subjective expert ratings \(3\.76 Aesthetic, 3\.88 Structure, and 42% Preference\)\. Beyond subjective ratings, the objective metric \(Category JSD\) uncovers the distinct decision\-making trade\-offs across strategies: \(1\)Beam Searchgreedily maximizes step\-wise aesthetic scores but lacks lookahead planning for delayed compositional rules, leading to the worst category distribution shift \(JSD=0\.0084\\text\{JSD\}=0\.0084\)\. \(2\)Rerankingrigidly enforces macro structural constraints post\-hoc, achieving the best Category JSD \(0\.00430\.0043\), but sacrifices overall stylistic harmony and local synergy \(lowest Human Pref\. at 26%\)\. \(3\)LE\-MCTSdynamically resolves this trade\-off via tree\-based lookahead search, effectively balancing immediate aesthetic compatibility with global structural constraints\.

Table 4:Ablation study on search strategies using the same model, candidate pool, and compute budget\. Five fashion experts are invited to evaluate from aesthetic, structure, and preference aspects\.Ablation on Search Budgets\.While Table[1](https://arxiv.org/html/2608.13888#S4.T1)provides the overall performance across different simulation budgetsMM, we further investigate how this hyperparameter influences the sequential composition dynamics\. In this experiment, the initial ensembleS0S\_\{0\}contains only one item sampled equally across six major fashion meta\-categories\. As illustrated in Figure[3](https://arxiv.org/html/2608.13888#S4.F3)\(a\), we report theOverall Score\(computed via the USCM value head\) and theΔ\\DeltaScore\(defined as the incremental gain between successive steps\), both averaged over all samples\. Additionally, the background gray bars indicate the average number of outfits at each step to illustrate the distribution of outfit lengths\.

The results highlight a clear advantage for look\-ahead planning\. During the critical early phases \(Steps 1 and 2\), LE\-MCTS\-100 \(Red Line\) and LE\-MCTS\-10 \(Purple Line\) achieve significantly higherΔ\\Deltascores than the greedy search \(M=1M=1\)\. While greedy search exhibits higherΔ\\Deltascores in the later phases, this potentially is a consequence of its lower score baseline caused by suboptimal early choices\. MCTS prioritizes long\-term compatibility, resulting in a significantly higher final overall score of5\.175\.17\. Furthermore, we observe a clear marginal effect regarding the number of simulations\. Performance scales consistently fromM=1M=1toM=100M=100\. Notably, LE\-MCTS\-10 achieves an inference latency of only0\.5​s0\.5\\text\{s\}per outfit \(compared to0\.08​s0\.08\\text\{s\}forM=1M=1and4\.7​s4\.7\\text\{s\}forM=100M=100\) while maintaining a performance profile remarkably close to LE\-MCTS\-100 \(final score5\.1465\.146vs\.5\.1725\.172\)\.M=10M=10establishes the optimal cost\-quality trade\-off for interactive, real\-time deployment\. In contrast, theType\-Awarebaseline \(black line\), which relies on stochastic sampling and greedy re\-ranking, lags severely with a final score of only1\.791\.79\. Its massive performance gap from Step 1 indicates a failure to capture high\-potential candidates without a learned policy prior\. This stark disparity confirms that purely discriminative models are insufficient for the CEG task; navigating the vast combinatorial space of fashion requires the synergistic integration of predictive priors and look\-ahead foresight\.

We observe that incremental gains for all MCTS variants peak at Step 2 and decline sharply thereafter, reflecting the inherent logic of fashion composition: once the core ensemble is established, the stylistic direction is largely defined, and subsequent additions like accessories provide diminishing marginal gains\. Since the bar chart indicates that most outfits contain22to44items, LE\-MCTS exerts its maximum influence exactly where the most impactful decisions are made\. Based on these findings, we recommend LE\-MCTS\-100 for high\-precision offline curation, while LE\-MCTS\-10 serves as a candidate for real\-time interactive systems\. The trajectory ofΔ\\Deltascores further validates the Marginal Gain\-based Terminal Operator described in Section[3\.5](https://arxiv.org/html/2608.13888#S3.SS5)\. By step 5, theΔ\\Deltascore for LE\-MCTS\-100 drops to0\.210\.21, approaching our predefined tolerance threshold ofϵ=0\.1\\epsilon=0\.1\. This convergence demonstrates the efficacy of the autonomous halting mechanism in identifying aesthetic saturation, preventing redundant or incompatible additions once a coherent ensemble is formed\.

![Refer to caption](https://arxiv.org/html/2608.13888v1/ablation_study.png)Figure 3:\(a\) Performance evolution across varying simulation budgetsMM\. Solid and dashed lines represent the overall compatibility score and incremental gain, respectively\. Gray bars denote the sample distribution at each composition step\. \(b\) Zero\-shot generalization performance on two unseen datasets: iFashion and PolyvoreU\. Bar charts represent the percentage score improvement over the greedy baseline, while lines indicate the incremental gain\.Ablation on Zero\-Shot Ensemble Generation\.To demonstrate the zero\-shot generalization of our framework, we conduct a cold\-start study on the unseeniFashionandPolyvoreUdatasets without any domain\-specific fine\-tuning\. As illustrated in Figure[3](https://arxiv.org/html/2608.13888#S4.F3)\(b\), the bars represent the percentage score improvement of LE\-MCTS \(M=100M=100\) over the Greedy \(M=1M=1\) baseline, while the lines denote the average incremental compatibility scores across five composition steps\. The results reveal highly synchronized performance trajectories across both unseen datasets\. At Step 1, LE\-MCTS achieves a substantial improvement of24\.75%24\.75\\%on iFashion and42\.17%42\.17\\%on PolyvoreU, demonstrating its efficacy in identifying critical stylistic anchors in cold\-start scenarios\. The incremental compatibility gains consistently peak at Step 2 across all domains \(Polyvore:1\.991\.99, iFashion:2\.112\.11, PolyvoreU:1\.911\.91\) before gradually converging\. Notably, the PolyvoreU trajectory terminates at Step 3, as the maximum ensemble length is set to four to reflect the dataset’s structural constraints \(i\.e\., a short average length of3\.353\.35due to the absence of accessories\)\. The highly synchronized performance trends across these distinct domains confirm that USCM captures universal, transferable stylistic heuristics rather than merely memorizing dataset\-specific co\-occurrences, thereby enabling robust generalization to evolving inventories within similar stylistic domains\. Visual examples of these zero\-shot generations across diverse seed categories are provided in Appendix[D\.3](https://arxiv.org/html/2608.13888#A4.SS3)\.

## 5Conclusion

In this paper, we address the inherent challenges of fashion outfit generation—namely, balancing explicit structural constraints with implicit aesthetic harmony within an exponentially large search space—by formalizing it as Constrained Ensemble Generation \(CEG\) framed as a deterministic Markov Decision Process\. To solve the CEG problem, we propose the Unified Sequential Composition Model \(USCM\), a core multi\-task representation architecture that jointly aligns continuous policy priors and compatibility value estimations within a shared multimodal manifold\. By coupling USCM with a Latent Expansion Monte Carlo Tree Search \(LE\-MCTS\) mechanism during generation, our framework actively navigates exponential candidate spaces without relying on hand\-crafted templates\. Extensive experiments on the Polyvore Outfits dataset, complemented by zero\-shot evaluations on the iFashion and PolyvoreU datasets, demonstrate the comprehensive superiority of our approach over existing paradigms across independent human preference evaluations, automated aesthetic proxies, and structural validity metrics\.

Despite these advancements, our framework presents several limitations\. First, fashion aesthetics are inherently highly subjective; thus, current evaluation protocols may not perfectly capture the complete spectrum of human preferences\. Second, although the USCM Value Head acts as an effective zero\-shot heuristic proxy to assess intermediate states, it is exclusively trained on complete outfits rather than explicitly optimized for partial ensemble evaluation\. Finally, while our approach reliably guarantees structurally sound and visually cohesive ensembles, it currently lacks the nuanced creativity of professional stylists and the sensitivity to adapt to specific occasional, regional, or seasonal contexts\. Future work will focus on addressing these gaps\.

## References

- \[1\]\(2024\)Fashion image\-to\-image translation for complementary item retrieval\.arXiv preprint arXiv:2408\.09847\.Cited by:[§1](https://arxiv.org/html/2608.13888#S1.p2.1),[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[2\]Y\. Aviv and A\. Pazgal\(2005\)A partially observed markov decision process for dynamic pricing\.Management science51\(9\),pp\.1400–1416\.Cited by:[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[3\]F\. Becattini, F\. M\. Teotini, and A\. Del Bimbo\(2023\)Transformer\-based graph neural networks for outfit generation\.IEEE Transactions on Emerging Topics in Computing12\(1\),pp\.213–223\.Cited by:[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[4\]C\. B\. Browne, E\. Powley, D\. Whitehouse, S\. M\. Lucas, P\. I\. Cowling, P\. Rohlfshagen, S\. Tavener, D\. Perez, S\. Samothrakis, and S\. Colton\(2012\)A survey of monte carlo tree search methods\.IEEE Transactions on Computational Intelligence and AI in games4\(1\),pp\.1–43\.Cited by:[§3\.4](https://arxiv.org/html/2608.13888#S3.SS4.SSS0.Px1.p1.1),[§3\.4](https://arxiv.org/html/2608.13888#S3.SS4.SSS0.Px2.p1.1)\.
- \[5\]C\. Chang, Y\. Chen, and D\. Jiang\(2025\)Using large multimodal models to predict outfit compatibility\.Decision Support Systems194,pp\.114457\.Cited by:[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[6\]W\. Chen, P\. Huang, J\. Xu, X\. Guo, C\. Guo, F\. Sun, C\. Li, A\. Pfadler, H\. Zhao, and B\. Zhao\(2019\)POG: personalized outfit generation for fashion recommendation at alibaba ifashion\.InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining,pp\.2662–2670\.Cited by:[§4\.1](https://arxiv.org/html/2608.13888#S4.SS1.p1.1)\.
- \[7\]P\. J\. Chia, G\. Attanasio, F\. Bianchi, S\. Terragni, A\. R\. Magalhães, D\. Goncalves, C\. Greco, and J\. Tagliabue\(2022\)Contrastive language and vision learning of general fashion concepts\.Scientific Reports12\(1\)\.External Links:ISSN 2045\-2322,[Document](https://dx.doi.org/10.1038/s41598-022-23052-9),[Link](https://doi.org/10.1038/s41598-022-23052-9)Cited by:[Appendix B](https://arxiv.org/html/2608.13888#A2.p1.1)\.
- \[8\]K\. Cui, S\. Liu, W\. Feng, X\. Deng, L\. Gao, M\. Cheng, H\. Lu, and L\. T\. Yang\(2025\)Correlation\-aware cross\-modal attention network for fashion compatibility modeling in ugc systems\.ACM Transactions on Multimedia Computing, Communications and Applications21\(11\),pp\.1–24\.Cited by:[§1](https://arxiv.org/html/2608.13888#S1.p2.1)\.
- \[9\]Z\. Cui, Z\. Li, S\. Wu, X\. Zhang, and L\. Wang\(2019\)Dressing as a whole: outfit compatibility learning based on node\-wise graph neural networks\.InThe world wide web conference,pp\.307–317\.Cited by:[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[10\]X\. Han, Z\. Wu, Y\. Jiang, and L\. S\. Davis\(2017\)Learning fashion compatibility with bidirectional lstms\.InProceedings of the 25th ACM international conference on Multimedia,pp\.1078–1086\.Cited by:[§1](https://arxiv.org/html/2608.13888#S1.p2.1),[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[11\]M\. Koren and O\. Peretz\(2026\)Dynamic colour dynamics: markov decision processes for fashion inventory management\.Annals of Operations Research358\(3\),pp\.1329–1359\.Cited by:[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[12\]J\. Li and Y\. Li\(2012\)Cognitive model based fashion style decision making\.Expert Systems with Applications39\(5\),pp\.4972–4977\.Cited by:[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[13\]T\. Lin, P\. Goyal, R\. Girshick, K\. He, and P\. Dollár\(2017\)Focal loss for dense object detection\.InProceedings of the IEEE international conference on computer vision,pp\.2980–2988\.Cited by:[§3\.3](https://arxiv.org/html/2608.13888#S3.SS3.SSS0.Px1.p1.1)\.
- \[14\]Y\. Lin, S\. Tran, and L\. S\. Davis\(2020\)Fashion outfit complementary item retrieval\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\.3311–3319\.Cited by:[§1](https://arxiv.org/html/2608.13888#S1.p2.1)\.
- \[15\]V\. Lipardi, D\. Dibenedetto, G\. Stamoulis, and M\. H\. Winands\(2025\)Quantum circuit design using a progressive widening enhanced monte carlo tree search\.Advanced Quantum Technologies8\(10\),pp\.e2500093\.Cited by:[§3\.4](https://arxiv.org/html/2608.13888#S3.SS4.SSS0.Px1.p1.1)\.
- \[16\]Z\. Lu, Y\. Hu, Y\. Jiang, Y\. Chen, and B\. Zeng\(2019\)Learning binary code for personalized fashion recommendation\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\.10562–10570\.Cited by:[§4\.1](https://arxiv.org/html/2608.13888#S4.SS1.p1.1)\.
- \[17\]A\. v\. d\. Oord, Y\. Li, and O\. Vinyals\(2018\)Representation learning with contrastive predictive coding\.arXiv preprint arXiv:1807\.03748\.Cited by:[§3\.3](https://arxiv.org/html/2608.13888#S3.SS3.SSS0.Px2.p1.1)\.
- \[18\]K\. Pang, X\. Zou, Z\. Broach, and W\. Wong\(2025\)Towards intelligent online cross\-selling\.Expert Systems with Applications,pp\.129686\.Cited by:[§1](https://arxiv.org/html/2608.13888#S1.p2.1)\.
- \[19\]S\. Saed and B\. Teimourpour\(2025\)Hybrid\-hierarchical fashion graph attention network for compatibility\-oriented and personalized outfit recommendation\.Machine Learning with Applications,pp\.100802\.Cited by:[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[20\]R\. Sarkar, N\. Bodla, M\. Vasileva, Y\. Lin, A\. Beniwal, A\. Lu, and G\. Medioni\(2022\)Outfittransformer: outfit representations for fashion recommendation\.InProceedings of the IEEE/CVF Conference on computer vision and pattern recognition,pp\.2263–2267\.Cited by:[§1](https://arxiv.org/html/2608.13888#S1.p2.1),[§1](https://arxiv.org/html/2608.13888#S1.p3.1),[§2](https://arxiv.org/html/2608.13888#S2.p1.1),[§4\.2](https://arxiv.org/html/2608.13888#S4.SS2.p1.1.5),[§4\.2](https://arxiv.org/html/2608.13888#S4.SS2.p2.1)\.
- \[21\]R\. Tan, M\. I\. Vasileva, K\. Saenko, and B\. A\. Plummer\(2019\)Learning similarity conditions without explicit supervision\.InProceedings of the IEEE/CVF international conference on computer vision,pp\.10373–10382\.Cited by:[§1](https://arxiv.org/html/2608.13888#S1.p2.1),[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[22\]M\. I\. Vasileva, B\. A\. Plummer, K\. Dusad, S\. Rajpal, R\. Kumar, and D\. Forsyth\(2018\)Learning type\-aware embeddings for fashion compatibility\.InProceedings of the European conference on computer vision \(ECCV\),pp\.390–405\.Cited by:[§2](https://arxiv.org/html/2608.13888#S2.p1.1),[§4\.1](https://arxiv.org/html/2608.13888#S4.SS1.p1.1),[§4\.2](https://arxiv.org/html/2608.13888#S4.SS2.p1.1)\.
- \[23\]Y\. Xu, W\. Wang, F\. Feng, Y\. Ma, J\. Zhang, and X\. He\(2024\)Diffusion models for generative outfit recommendation\.InProceedings of the 47th international ACM SIGIR conference on research and development in information retrieval,pp\.1350–1359\.Cited by:[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[24\]M\. Yu, Y\. Ma, L\. Wu, C\. Wang, X\. Li, and L\. Meng\(2025\)FashionDPO: fine\-tune fashion outfit generation model using direct preference optimization\.InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval,pp\.212–222\.Cited by:[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.
- \[25\]D\. Zhou, H\. Zhang, K\. Yang, L\. Liu, H\. Yan, X\. Xu, Z\. Zhang, and S\. Yan\(2022\)Learning to synthesize compatible fashion items using semantic alignment and collocation classification: an outfit generation framework\.IEEE Transactions on Neural Networks and Learning Systems35\(4\),pp\.5226–5240\.Cited by:[§2](https://arxiv.org/html/2608.13888#S2.p1.1)\.

## Appendix ALE\-MCTS Algorithmic Details

Algorithm 1Latent Expansion Monte Carlo Tree Search \(LE\-MCTS\)1:Input:Root state

S0S\_\{0\}, USCM model

fθf\_\{\\theta\}, VectorDB

𝒟\\mathcal\{D\}, Total simulations

MM
2:Hyperparameters:

Cp​u​c​t,Cp​w,αp​w,τ,kC\_\{puct\},C\_\{pw\},\\alpha\_\{pw\},\\tau,k
3:root

←\\leftarrowCreateNode\(

S0S\_\{0\}\)⊳\\trianglerightInitialize the search tree withS0S\_\{0\}

4:for

m=1m=1to

MMdo

5:node

←\\leftarrowroot

6:// Phase 1: Selection with Progressive Widening

7:whilenode\.

i​s​\_​l​e​a​fis\\\_leafis Falsedo

8:ifnode\.

i​s​\_​i​n​i​t​i​a​l​i​z​e​dis\\\_initializedis Falsethen⊳\\trianglerightInitialize Current Node

9:

vn​o​d​e,𝐳^n​o​d​e←fθ\(node\.S\)v\_\{node\},\\hat\{\\mathbf\{z\}\}\_\{node\}\\leftarrow f\_\{\\theta\}\(node\.S\)
10:node\.

c​a​n​d​i​d​a​t​e​s←𝒟\.s​e​a​r​c​h​\(𝐳^n​o​d​e,k\)candidates\\leftarrow\\mathcal\{D\}\.search\(\\hat\{\\mathbf\{z\}\}\_\{node\},k\)
11:

n​o​d​e\.p←softmax​\(s​i​m/τ\)node\.\\textbf\{p\}\\leftarrow\\text\{softmax\}\(sim/\\tau\)⊳\\trianglerightSet prior probabilities for candidates

12:node\.

i​s​\_​i​n​i​t​i​a​l​i​z​e​dis\\\_initialized←\\leftarrowTrue

13:endif

14:if

\|\|node

\.children\|<\(Cp​w⋅nn​o​d​e\)αp​w\.children\|<\(C\_\{pw\}\\cdot n\_\{node\}\)^\{\\alpha\_\{pw\}\}andnode\.

c​a​n​d​i​d​a​t​e​scandidatesis not emptythen

15:break⊳\\trianglerightTrigger Expansion

16:else

17:node

←arg⁡maxc∈node\.c​h​i​l​d​r​e​n⁡\(qc\+Cp​u​c​t⋅pc​nn​o​d​e1\+nc\)\\leftarrow\\arg\\max\_\{c\\in\\text\{node\}\.children\}\\left\(q\_\{c\}\+C\_\{puct\}\\cdot p\_\{c\}\\frac\{\\sqrt\{n\_\{node\}\}\}\{1\+n\_\{c\}\}\\right\)⊳\\trianglerightchild\.q is the mean value

18:endif

19:endwhile

20:// Phase 2: Expansion & Evaluation

21:

a←a\\leftarrownode\.

c​a​n​d​i​d​a​t​e​s\.p​o​p​\(0\)candidates\.pop\(0\)⊳\\trianglerightSelect the best candidate from the pool

22:

S′←n​o​d​e\.S∪\{a\}S^\{\\prime\}\\leftarrow node\.S\\cup\\\{a\\\}⊳\\trianglerightState transition: append new item

23:

v′,\_←fθ​\(S′\)v^\{\\prime\},\\\_\\leftarrow f\_\{\\theta\}\(S^\{\\prime\}\)⊳\\trianglerightEvaluate the new state

24:child

←C​r​e​a​t​e​N​o​d​e​\(S′,parent=n​o​d​e\)\\leftarrow CreateNode\(S^\{\\prime\},\\text\{parent\}=node\)
25:

n​o​d​e\.c​h​i​l​d​r​e​n←n​o​d​e\.c​h​i​l​d​r​e​n∪\{c​h​i​l​d\}node\.children\\leftarrow node\.children\\cup\\\{child\\\}
26:

n​o​d​e←c​h​i​l​dnode\\leftarrow child⊳\\trianglerightPrepare for backpropagation from the new leaf

27:// Phase 3: Backpropagation

28:while

n​o​d​enodeis notNulldo

29:

n​o​d​e\.n←n​o​d​e\.n\+1node\.n\\leftarrow node\.n\+1
30:

n​o​d​e\.w←n​o​d​e\.w\+v′node\.w\\leftarrow node\.w\+v^\{\\prime\}
31:

n​o​d​e\.q←n​o​d​e\.w/n​o​d​e\.nnode\.q\\leftarrow node\.w/node\.n
32:

n​o​d​e←n​o​d​e\.p​a​r​e​n​tnode\\leftarrow node\.parent
33:endwhile

34:endfor

35:Return

a∗=arg⁡maxc∈r​o​o​t\.c​h​i​l​d​r​e​n⁡nca^\{\*\}=\\arg\\max\_\{c\\in root\.children\}n\_\{c\}⊳\\trianglerightSelect action with most visits

## Appendix BImplementation Details

USCM Details\.For the USCM architecture, we employ a 6\-layer Transformer encoder with 16 attention heads and a feed\-forward dimension of 2024\. We use the pre\-trained FashionCLIP\([7](https://arxiv.org/html/2608.13888#bib.bib6)\)as the item encoder to extract 512\-dimensional visual features only\. The model is trained for 50,000 steps with a batch size of 1,000 on the Polyvore Outfits dataset\. We use the AdamW optimizer with a peak learning rate of2×10−52\\times 10^\{\-5\}\. To ensure stable convergence, we adopt the OneCycleLR scheduler with a cosine annealing strategy, where the warm\-up phase accounts for 30% of the total steps\. Regarding the loss functions, the temperature parameterτ\\taufor InfoNCE loss is set to 0\.2, and the NIP loss weightλp\\lambda\_\{p\}is 0\.1\.

All training experiments for the USCM were conducted on a single NVIDIA GeForce RTX 4090 D GPU \(24GB VRAM\)\. The total training duration for the 50,000\-iteration schedule was approximately 8 hours and 20 minutes\. During training, the model consumed a peak GPU memory of approximately 12\.5 GB\.

LE\-MCTS Details\.We set the number of simulationsM=100M=100and the candidate retrieval breadthk=10k=10\. For progressive widening, the hyperparameters are set toCp​w=1\.0C\_\{pw\}=1\.0andαp​w=0\.5\\alpha\_\{pw\}=0\.5\. The exploration constantCp​u​c​tC\_\{puct\}is fixed at 40\.0 to balance the semantic prior and search value\. The search process automatically terminates if the marginal value gain falls below the thresholdδ=0\.1\\delta=0\.1\. For prior probability calculation, the temperatureτ\\tauis set to 1\.0\.

CEG Experiment Details\.To align with the average outfit size of 5\.34 observed in the Polyvore dataset, we limit the maximum ensemble length to six items for all methods\. The initial ensembleS0S\_\{0\}is constructed by randomly sampling one or two items from core categories \(e\.g\., tops, bottoms, shoes, all\-body, and outerwear\) within ground\-truth outfits\. For theVLLMbaseline, we utilize thegemini\-3\.0\-flashmodel by simultaneously providing the initial ensemble’s images as visual prompts alongside the text prompt:“You are a professional fashion stylist\. Based on the images provided by the user, please recommend five items that would enhance the overall harmony of the outfit\. Your answer must be in English\. Start item’s description is\{description\}”\. The generated textual descriptions are subsequently parsed, encoded using FashionCLIP, and matched against the candidate pool via Top\-1 vector similarity search\.

## Appendix CDetailed Definitions of Structural Metrics

To rigorously assess the structural integrity of the generated ensembles, we provide the formal definitions for theCategory Jensen\-Shannon Divergence \(Cate\. JSD\)and theEnsemble Completeness ScoreSfullS\_\{\\text\{full\}\}\.

### C\.1Category Jensen\-Shannon Divergence \(Cate\. JSD\)

The Cate\. JSD quantifies the similarity between the generated category distribution and the dataset’s natural distribution\. Let𝒴\\mathcal\{Y\}be the set ofKKunique fashion categories\. We define the empirical ground\-truth distributionPPand the generated distributionQQas probability vectors inℝK\\mathbb\{R\}^\{K\}, where each element represents the frequency of a category\. The JSD is defined as:

JSD\(P∥Q\)=12DK​L\(P∥M\)\+12DK​L\(Q∥M\)\\text\{JSD\}\(P\\parallel Q\)=\\frac\{1\}\{2\}D\_\{KL\}\(P\\parallel M\)\+\\frac\{1\}\{2\}D\_\{KL\}\(Q\\parallel M\)\(2\)whereM=12​\(P\+Q\)M=\\frac\{1\}\{2\}\(P\+Q\)is the average distribution, andDK​LD\_\{KL\}is the Kullback\-Leibler divergence\. Using Shannon entropyH\(P\)=−∑i=1Kpilog2piH\(P\)=\-\\sum\_\{i=1\}^\{K\}p\_\{i\}\\log\_\{2\}p\_\{i\}, the JSD can be computed as:

JSD\(P∥Q\)=H\(M\)−12\(H\(P\)\+H\(Q\)\)\\text\{JSD\}\(P\\parallel Q\)=H\(M\)\-\\frac\{1\}\{2\}\\big\(H\(P\)\+H\(Q\)\\big\)\(3\)
A value of00indicates identical distributions\. This metric penalizes models that over\-generatesafecategories \(e\.g\., accessories\) to inflate compatibility scores\.

### C\.2Ensemble Validity Score \(SvalidS\_\{\\text\{valid\}\}\)

While we state in the main text that perfect categorical grammar cannot be exhaustively hand\-coded as explicit generative constraints, a simplified heuristic rule remains a practical and highly effective approximate metric for benchmarking structural integrity\. A coherent outfit must satisfy basic functional requirements and logical consistency for human wear\. We define a validity functionfvalid​\(S\)∈\{0,1\}f\_\{\\text\{valid\}\}\(S\)\\in\\\{0,1\\\}for an ensembleSS\. LetCats​\(S\)\\text\{Cats\}\(S\)be the set of categories present inSS\. Given the specific category taxonomy of the Polyvore Outfits dataset, we stipulate that an outfit isvalidonly if it satisfies bothfunctional coverageandlogical compatibility:

1. 1\.Functional Coverage:SSmust satisfy at least one of the following core combinations: - •\{tops,bottoms,shoes\}⊆Cats​\(S\)\\\{\\text\{tops\},\\text\{bottoms\},\\text\{shoes\}\\\}\\subseteq\\text\{Cats\}\(S\) - •\{all\-body,shoes\}⊆Cats​\(S\)\\\{\\text\{all\-body\},\\text\{shoes\}\\\}\\subseteq\\text\{Cats\}\(S\)
2. 2\.Logical Compatibility:To prevent categorical conflicts, an outfit containing an all\-body item must not simultaneously include tops or bottoms\.

TheSvalidS\_\{\\text\{valid\}\}score for a test set ofNNgenerated outfits is defined as the average success rate:

Svalid=1N​∑i=1Nfvalid​\(Si\)S\_\{\\text\{valid\}\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}f\_\{\\text\{valid\}\}\(S\_\{i\}\)\(4\)This metric evaluates whether the framework generates functionally complete and logically sound ensembles, rather than merely aggregating items\. To empirically validate this approximate rule, we observe that 78\.38% of the human\-authored ground\-truth outfits in the dataset naturally satisfy these criteria, demonstrating the robustness and practical relevance of utilizing this heuristic as an evaluative baseline\.

## Appendix DQualitative Results

### D\.1Extensive Qualitative Comparisons

![Refer to caption](https://arxiv.org/html/2608.13888v1/qualitative_results_compare.png)Figure 4:Qualitative comparison of outfit generation starting from different seed item categories: \(a\) tops, \(b\) bottoms, \(c\) shoes, and \(d\) all\-body\. Our method is compared against OutfitTransformer, Type\-aware, and the original Human Composed reference outfits\.To further demonstrate the robustness and versatility of our framework, we provide an extensive qualitative comparison in Figure[4](https://arxiv.org/html/2608.13888#A4.F4)\. We compare our generated ensembles against two state\-of\-the\-art baselines, OutfitTransformer and Type\-aware, alongside the original Human Composed outfits from the dataset\. The generations are evaluated under four distinct initialization scenarios based on the category of the seed item: starting from \(a\) tops, \(b\) bottoms, \(c\) shoes, and \(d\) all\-body items\.

As illustrated, our method consistently generates outfits with superior aesthetic coordination and more reasonable category coverage compared to the baselines\. One critical phenomenon is worth noting regarding the dataset quality\. As shown in theHuman Composedcolumns, the ground\-truth outfits frequently exhibit uncoordinated stylistic choices and redundant category overlaps\. This clearly indicates that the underlying training dataset contains a significant amount of noise\. Despite learning from such noisy and imperfect supervision, our method demonstrates a superior ability to consistently synthesize more aesthetically harmonious and structurally coherent outfits than the baseline methods\.

### D\.2Qualitative Analysis: Generating Variations from Same Seed

In this section, we explore the model’s capacity forvariation outfit generation\. While the standard LE\-MCTS framework \(as described in Section[3\.4](https://arxiv.org/html/2608.13888#S3.SS4)\) utilizes a deterministic selection rule—picking the action corresponding to the child node with the maximum visit countnnto ensure the most robust decision—we relax this constraint during this qualitative study to observe the richness of the search space\.

Specifically, we transform the deterministic selection into a stochastic sampling process\. For a given stateStS\_\{t\}, after the simulation budgetMMis exhausted, we obtain the visit countsnin\_\{i\}for all expanded child nodescic\_\{i\}\. Instead of applyingarg⁡max⁡\(ni\)\\arg\\max\(n\_\{i\}\), we compute a selection probability distribution via a Softmax operator:

P⁡\(ai\|St\)=exp⁡\(ni/τ\)∑jexp⁡\(nj/τ\)P\(a\_\{i\}\|S\_\{t\}\)=\\frac\{\\exp\(n\_\{i\}/\\tau\)\}\{\\sum\_\{j\}\\exp\(n\_\{j\}/\\tau\)\}whereτ\\tauis a temperature parameter\. By sampling from this distribution, the agent is able to follow different high\-potential branches that were explored during the MCTS phase\.

![Refer to caption](https://arxiv.org/html/2608.13888v1/qualitative_results_variation.png)Figure 5:Variation outfit generation via stochastic latent sampling\. Each row represents a unique generation session starting from the same initial seed item\.As illustrated in Figure[5](https://arxiv.org/html/2608.13888#A4.F5), starting from identical seed items, the framework successfully curates a wide array of variations that differ in silhouette, item category, and texture while maintaining a consistent global harmony\. For instance, given a brown knit dress as a seed, the model explores different paths: one focusing on a formal aesthetic with long coats and leather boots, while another leans toward a casual style with lighter cardigans and suede accessories\. The ability to generate such high\-quality variations underscores the superiority of our approach\.

### D\.3Qualitative Analysis: Generalization to Unseen Datasets

To evaluate the zero\-shot transferability of our framework, we visualize generation results on two unseen datasets, iFashion and PolyvoreU, in Figure[6](https://arxiv.org/html/2608.13888#A4.F6)\. Each column corresponds to a fixed category for the initial seed item to test the framework’s consistency across diverse starting points\.

As shown in Figure[6](https://arxiv.org/html/2608.13888#A4.F6)\(a\), our method effectively adapts to the distinct style distribution of the iFashion dataset\. For the PolyvoreU dataset in Figure[6](https://arxiv.org/html/2608.13888#A4.F6)\(b\), we impose a maximum ensemble length ofT=4T=4\. This constraint is specifically designed to align with the dataset’s underlying distribution, which has an average outfit length of3\.353\.35and lacks accessory items\. Despite these structural differences and the absence of fine\-tuning on these specific domains, our LE\-MCTS planner successfully navigates the new latent landscapes to synthesize stylistically coherent outfits, demonstrating the strong generalization capability of the learned USCM heuristics\.

![Refer to caption](https://arxiv.org/html/2608.13888v1/qualitative_results_other_datasets.png)Figure 6:Zero\-shot generation results on the iFashion and PolyvoreU datasets\. Columns are organized by the category of the initial seed item\. For PolyvoreU, the maximum ensemble length is set to 4 to reflect the dataset’s characteristics\.