ACM ICMR 2026 · Amsterdam · DOI: 10.1145/3805622.3810590

VietFashion: Benchmarking Sketch–Text
Composed Image Retrieval for Cultural Outfits

A benchmark for fine-grained cultural garment retrieval centered on the Vietnamese Áo Dài

Hoang-Nguyen Cao *
Le-Hoang Bui *
Dinh-Khoi Vo
Minh-Triet Tran
Trung-Nghia Le

* Equal contribution    Corresponding author

University of Science & Vietnam National University, Ho Chi Minh City, Vietnam

ICMR '26, June 16–19, 2026, Amsterdam, Netherlands  ·  https://doi.org/10.1145/3805622.3810590

Paper Code Project Page
Representative sketch-photo pairs

Figure 1: Representative sketch-photo pairs of Áo Dài from our VietFashion dataset. The top row features sketches reflecting diverse levels of abstraction and detail, while the bottom rows present corresponding photographs of authentic garments.

650
Human-drawn sketches
21K
Synthesized garment images
7,000
Composed retrieval queries
1→3
Multi-target mapping
11
Attribute categories

Abstract

Cultural garments pose a unique challenge for visual retrieval systems, as their identity often depends on subtle structural and symbolic details that are poorly captured by standard AI models. We introduce VietFashion, a new benchmark for sketch–text composed image retrieval centered on the Áo Dài, a traditional Vietnamese garment. VietFashion enables designers and researchers to retrieve culturally meaningful outfits using a combination of hand-drawn sketches — which convey garment structure — and textual descriptions — which encode cultural semantics.


The dataset is initialized with 650 sketches and expanded using generative models to produce over 21,000 photorealistic images with aligned captions. Textual prompts describing detailed outfit attributes are extracted from fashion magazines to ensure authenticity and diversity. To better reflect the inherent ambiguity of design intent, VietFashion adopts a multi-target retrieval setting, where a single query may correspond to multiple valid results. Experimental results reveal significant performance gaps in modeling fine-grained cultural semantics and multi-modal composition, positioning VietFashion as a challenging benchmark for fine-grained fashion retrieval.

Dataset available at: https://hng0303.github.io/VietFashion

Dataset Construction Pipeline

A two-stage generative pipeline produces aligned sketch–text–image triplets from curated cultural attributes.

Pipeline overview

Figure 2: Overview of the VietFashion dataset construction pipeline. The pipeline begins with sketches (S) and sampled garment attributes (A). We utilize SANA-ControlNet to generate multi-target images (I) under spatial constraints, while Qwen-2.5-Instruct distills attributes into concise natural language captions (C) to form the final composed retrieval triplet.

STAGE 0
Sketch Collection
650 human-drawn sketches covering diverse Áo Dài silhouettes, collar types, sleeve variants, and structural compositions. Balanced across abstraction levels (low/medium/high).
650 sketches
STAGE 1
Attribute-Driven Synthesis
Random sampling from 11 curated attribute categories (fabric, neckline, sleeves, embroidery…). SANA-ControlNet generates photorealistic targets conditioned on sketch + structured prompt.
SANA-ControlNet
STAGE 2
Caption Refinement
Qwen-2.5 3B Instruct distills structured attributes into concise, neutral single-sentence captions starting with "A photo of…" (avg. 42.46 words).
Qwen-2.5 3B
STAGE 3
Triplet Alignment
Each query (Sketch, Caption) is paired with 3 semantically consistent target images, reducing false-negative supervision inherent in single-target designs.
1 → 3 mapping

Multi-Target Query Design

In real-world fashion scenarios, multiple garments may satisfy the same semantic description. Using only one positive target during training causes valid alternatives to be misclassified as negatives — a false-negative supervision problem. VietFashion addresses this by adopting a 1→3 multi-target mapping: each query is paired with three semantically consistent but visually distinct target images.

Dataset examples

Figure 3: Examples from the proposed VietFashion dataset. Each query contains a sketch of an Áo Dài, a natural-language caption describing garment attributes and context, and multiple valid target images.

Query format:   (Sketch, Caption)  →  {Target Img.1, Target Img.2, Target Img.3}

Data splits:   5,200 training queries  ·  650 validation queries  ·  1,150 testing queries   (split at sketch level to prevent query-level leakage)

Comparison with Existing Datasets

VietFashion uniquely targets cultural outfits and employs multi-target supervision to address the ambiguity inherent in fine-grained sketch-text retrieval.

DatasetYearDomainQuery ModalityCIRMulti-Target
TU-Berlin2012General ObjectSketch
Sketchy Extended2016General ObjectSketch
FashionIQ2019Western FashionImage, Text
QuickDraw-Ext2019General ObjectSketch
CIRR2021Open-DomainImage, Text
CIRCO2022Open-DomainImage, Text
FACap2025FashionImage, Text
CSTBIR2025General ObjectSketch, Text
FIGROTD2026General ObjectSketch, Text, Image
VietFashion (Ours)2026Cultural OutfitSketch, Text

Benchmark Results

Retrieval performance on the VietFashion test set. Red = best per column. Methods are categorized by learning paradigm.

MethodParadigmR@1R@5R@10mAPMRR
ZSE-SBIRSBIR0.02850.06230.10770.03230.0539
S3BIR-DINOSBIR0.01570.05650.09480.02160.0428
TaskFormerST-CIR0.05640.14720.20670.02690.0891
VaGFeMST-CIR0.07500.16120.22010.03560.1142
CLIP4CIRSupervised0.03130.11490.18510.10640.1908
BLIP4CIRSupervised0.08770.26720.37030.24830.3950
SEARLE-ViT/BZero-shot0.00000.02000.04000.02000.0500
SEARLE-ViT/LZero-shot0.01000.03000.04000.03000.0600
Pic2WordZero-shot0.00820.02100.03640.02530.0523
Pic2Word (Fine-tuned)Fine-tuned0.00870.02210.03740.02530.0527

Pic2Word (Fine-tuned) was adapted using sketches in our training set.

Key Findings

Multimodal composition matters

VaGFeM achieves ~2.6× higher R@1 than the best SBIR baseline, confirming that textual attribute conditioning is essential for fine-grained cultural retrieval.

Zero-shot models struggle

Best zero-shot model reaches only R@1 = 0.01. General vision-language pretraining doesn't transfer to abstract sketches paired with cultural garment semantics.

Fine-grained retrieval is hard

Even the best model achieves R@1 below 0.09 while R@10 reaches 0.37 — many Áo Dài share nearly identical silhouettes differing only in subtle embroidery or collar details.

Architecture sensitivity

BLIP4CIR outperforms CLIP4CIR by a large margin (MRR 0.395 vs 0.191), suggesting stronger text-visual grounding is critical when visual differences are semantically subtle.

Multi-target complexity

Low mAP across all methods indicates the 1→3 design introduces genuine ambiguity that requires attribute-level discrimination rather than instance memorization.

Caption complexity trade-off

Captions average 42.46 words, ensuring cultural richness. The BLIP4CIR vs CLIP4CIR gap suggests models with stronger fine-grained grounding better parse these long descriptors.

Contributions

Citation

If you find VietFashion useful in your research, please cite our paper:

@inproceedings{cao2026vietfashion,
  author = {Hoang-Nguyen Cao and Le-Hoang Bui and Dinh-Khoi Vo
            and Minh-Triet Tran and Trung-Nghia Le},
  title = {VietFashion: Benchmarking Sketch–Text Composed Image
            Retrieval for Cultural Outfits},
  booktitle = {International Conference on Multimedia Retrieval
            (ICMR '26)},
  year = {2026},
  address = {Amsterdam, Netherlands},
  doi = {10.1145/3805622.3810590},
}
Acknowledgments
This research is funded by Vietnam National University – Ho Chi Minh City (VNU-HCM) under Grant Number B2026-18-17.