Garments2Look [CVPR 2026]

Paper: Garments2Look: A Multi-Reference Dataset for High-Fidelity Outfit-Level Virtual Try-On with Clothing and Accessories

Authors: Junyao Hu, Zhongwei Cheng, Waikeung Wong, Xingxing Zou

g2l.mp4

News and updates

[2026-10-11] Released final Qwen-Image-2.1 inpainting and editing LoRAs, with both-model inference and paired seed-0 examples.
[2026-10-04] Updated the dataset to 98,012 outfit records, with five annotation types, refined and dilated v3 masks, OOTD collages, and editing source images.
[2026-10-04] Released Qwen-Image-Edit-2509 LoRAs for inpainting and editing.
[2026-04-09] Garments2Look was accepted to CVPR 2026.
[2026-03-17] Released the Garments2Look dataset, including all image data and inputs for the inpainting task setting.
[2026-03-14] Submitted the first version of our paper to arXiv.

TODO

Release Qwen 2509 LoRAs for inpainting and editing.
Add dataset preparation, training, inference, and examples.
Release editing-task inputs, v1.1 outfit annotations, and improved v3 masks.
Train and open-source Qwen Image 2.1 LoRAs.

Overview

Virtual try-on (VTON) has advanced single-garment visualization, yet real-world fashion centers on full outfits with multiple garments, accessories, fine-grained categories, layering, and diverse styling, remaining beyond current VTON systems. Existing datasets are category-limited and lack outfit diversity. We introduce Garments2Look, the first large-scale multimodal dataset for outfit-level VTON, comprising approximately 80K many-garments-to-one-look pairs in the paper, expanded to 98,012 indexed outfit records in the current dataset across 40 major categories and 300+ fine-grained subcategories. Each pair includes an outfit with 3-12 reference garment images (4.48 items per outfit on average), a model image wearing the outfit, and detailed item and try-on textual annotations. To balance authenticity and diversity, we propose a synthesis pipeline. It involves heuristically constructing outfit lists before generating try-on results, with the entire process subjected to strict automated filtering and human validation to ensure data quality. To probe task difficulty, we adapt SOTA VTON methods and general-purpose image editing models to establish baselines. Results show current methods struggle to try on complete outfits seamlessly and to infer correct layering and styling, leading to misalignment and artifacts.

Dataset comparison. Garments2Look targets complete outfits with multiple clothing items and accessories, including annotations of layering and styling.

Dataset construction

Overview of Garments2Look construction process. Our dataset follows four steps: (1) Data Collection: obtaining real-world clothing items and their outfit suggestions from different sources; (2) Data Synthesis: enriching the dataset content and diversity by generating new outfit lists and look images; (3) Data Filtering: ensuring visual consistency and data quality, including annotations of garment images, outfit lists and look images; and (4) Data Evaluation: verifying the data quality, designing new metrics for outfit-level VTON task, and testing SOTA models.

Sample outfits and annotations. Each sample pairs multiple reference items with a look image and annotations describing the outfit, layering, and styling.

Qwen LoRA models

Checkpoints

Base model	Task	File	Training
Qwen-Image-Edit-2509	Inpainting	`Qwen-Image-Edit-2509-LoRA-2-refer-20k-inpainting-epoch-1.safetensors`	20,000 samples; 2 completed epochs; rank 32
Qwen-Image-Edit-2509	Editing	`Qwen-Image-Edit-2509-LoRA-2-refer-20k-editing-epoch-1.safetensors`	20,000 samples; 2 completed epochs; rank 32
Qwen-Image-2.1	Inpainting	`Qwen-Image-2.1-LoRA-2-refer-97068-inpainting-epoch-0.safetensors`	97,068 samples; 1 completed epoch; final step 48,534; rank 32
Qwen-Image-2.1	Editing	`Qwen-Image-2.1-LoRA-2-refer-97068-editing-epoch-0.safetensors`	97,068 samples; 1 completed epoch; final step 48,534; rank 32

Epoch numbers are zero-indexed: epoch-1 means two completed epochs, and epoch-0 means one completed epoch. Both 2.1 files are the final step-48534 checkpoints, not step-48000. These are adapters, not full models. Use the matching base model and task. File checksums and training metadata are in HF manifest.json.

Installation and download

Tested with Python 3.10, PyTorch 2.7.1 / CUDA 12.8, Transformers 5.16.1, and H200 GPUs. Run commands from the repository root. Clone the code repository first:

git clone https://github.com/ArtmeScienceLab/Garments2Look.git
cd Garments2Look
``` Download only the base model you intend to use.

```bash
conda create -n g2l-lora python=3.10 -y
conda activate g2l-lora
python -m pip install torch==2.7.1 torchvision==0.22.1 --index-url https://download.pytorch.org/whl/cu128
python -m pip install huggingface_hub
hf download ArtmeScienceLab/Garments2Look-LoRA --local-dir models/Garments2Look-LoRA
python -m zipfile -e models/Garments2Look-LoRA/qwen-runtime.zip models/Garments2Look-LoRA
python -m pip install -e models/Garments2Look-LoRA/qwen-runtime
python -m pip install transformers==5.16.1

hf download Qwen/Qwen-Image-Edit-2509 --local-dir models/Qwen-Image-Edit-2509
hf download Qwen/Qwen-Image-2.1 --local-dir models/Qwen-Image-2.1

The repository inference script imports the extracted runtime. Adapters use the DiffSynth format; compatibility with other LoRA loaders has not been verified. Set CUDA_VISIBLE_DEVICES to choose a GPU.

Inpainting inference

Qwen-Image-Edit-2509

python scripts/inference/inference.py --model-version 2509 --task inpainting \
  --model-dir models/Qwen-Image-Edit-2509 \
  --lora models/Garments2Look-LoRA/Qwen-Image-Edit-2509-LoRA-2-refer-20k-inpainting-epoch-1.safetensors \
  --origin examples/qwen2509/input-inpainting.png \
  --ootd examples/qwen2509/ootd.png \
  --prompt-file examples/qwen2509/prompt-inpainting.txt \
  --output output/qwen2509-inpainting.png --seed 0 --steps 40 --cfg-scale 4

Qwen-Image-2.1

python scripts/inference/inference.py --model-version 2.1 --task inpainting \
  --model-dir models/Qwen-Image-2.1 \
  --lora models/Garments2Look-LoRA/Qwen-Image-2.1-LoRA-2-refer-97068-inpainting-epoch-0.safetensors \
  --origin examples/qwen21/input-inpainting.png \
  --ootd examples/qwen21/ootd.png \
  --prompt-file examples/qwen21/prompt-inpainting.txt \
  --output output/qwen21-inpainting.png --seed 0 --steps 40 --cfg-scale 1

Editing inference

Qwen-Image-Edit-2509

python scripts/inference/inference.py --model-version 2509 --task editing \
  --model-dir models/Qwen-Image-Edit-2509 \
  --lora models/Garments2Look-LoRA/Qwen-Image-Edit-2509-LoRA-2-refer-20k-editing-epoch-1.safetensors \
  --origin examples/qwen2509/input-editing.png \
  --ootd examples/qwen2509/ootd.png \
  --prompt-file examples/qwen2509/prompt-editing.txt \
  --output output/qwen2509-editing.png --seed 0 --steps 40 --cfg-scale 4

Qwen-Image-2.1

python scripts/inference/inference.py --model-version 2.1 --task editing \
  --model-dir models/Qwen-Image-2.1 \
  --lora models/Garments2Look-LoRA/Qwen-Image-2.1-LoRA-2-refer-97068-editing-epoch-0.safetensors \
  --origin examples/qwen21/input-editing.png \
  --ootd examples/qwen21/ootd.png \
  --prompt-file examples/qwen21/prompt-editing.txt \
  --output output/qwen21-editing.png --seed 0 --steps 40 --cfg-scale 1

Default seed is 0, with 40 steps. If CFG is omitted, the script uses 4 for 2509 and 1 for 2.1. Images use a 1,048,576-pixel budget, aligned to multiples of 16 for 2509 and 32 for 2.1. Each invocation saves its PNG and a JSON parameter record. For custom outfits, replace Figure 1, the collage, and the full prompt.

Paired example: P00958796_b1

This five-item test outfit includes a top, jacket, shorts, sandals, and bag. Styling specifies a tucked-in top and an unbuttoned jacket; layering is top → jacket. Both task prompts are supplied in the example folders.

Qwen-Image-Edit-2509 LoRA

Qwen-Image-2.1 LoRA

Columns: OOTD, inpainting input, inpainting output, editing input, editing output. Full task prompts are printed below each figure. Both models use seed 0 and 40 steps, with CFG 4 for 2509 and CFG 1 for 2.1. The 2.1 images use the final adapters released above.

Full prompt (both tasks):

Keep the woman's identity, pose, background in Figure 1 unchanged, wearing the outfit in Figure 2, include (1) a top (tucked-in), (2) a jacket (unbuttoned), (3) shorts, (4) sandals, (5) a bag. Layering Order: (1) -> (2).

This example was selected after comparing six candidate outfits at fixed seed 0, four outputs per candidate. It illustrates a selected successful case, not aggregate test performance. The two base models were trained on different data counts and epoch counts. Bag shape, garment hems, and pose may still differ. Selection provenance is in examples/selection.json; each output has a parameter JSON, and target images are included.

Training details

Setting	2509	2.1
Training examples	20,000	Full 97,068 training split
Completed epochs	2	1
LoRA rank	32	32
Learning rate	1e-4	1e-4
Pixel budget	1,048,576	1,048,576
Precision	BF16	BF16
Gradient checkpointing	Enabled	Enabled
Final step	Epoch-based checkpoint	48,534

The 2.1 runs used two H200 GPUs and global batch size 2. Training took approximately 58h36m (inpainting) and 58h50m (editing). These adapters are later releases than the paper's rebuttal-stage models; the selected examples do not establish aggregate improvements over those models.

For dataset preparation and the public 2509 training workflow, see the code repository. The HF runtime and this repository's inference script support both models. The unified training entry point supports both models. The task-specific 2.1 launchers below use the configuration of the released training runs.

Dataset preparation and training

Download training dataset

Skip this section for inference with the bundled examples or your own prepared input images. Full dataset downloads are needed for training and dataset analysis.

For training or dataset analysis, download the annotations and image archives from Hugging Face:

python -m pip install -U huggingface_hub
hf download ArtmeScienceLab/Garments2Look --repo-type dataset \
  --local-dir /path/to/Garments2Look-data

The current release contains 56 independent .tar.gz archives, approximately 437.54 GB (407.49 GiB) compressed, plus JSON annotations (approximately 438.1 GB in total). Each archive includes paths beginning with mytheresa/ or polyvore/. Extract every archive into the dataset root:

cd /path/to/Garments2Look-data
sha256sum -c SHA256SUMS
find mytheresa polyvore -type f -name '*.tar.gz' -print0 |
  while IFS= read -r -d '' archive; do
    tar -xzf "$archive" -C .
  done

Each shard can be extracted independently. Keep the original subset layout:

Garments2Look-data/
├── Garments2Look.py
├── mytheresa_image_v1.0_2512.json
├── mytheresa_outfit_v1.0_2512.json
├── mytheresa_outfit_v1.1_2512.json
├── polyvore_image_v1.0_2512.json
├── polyvore_outfit_v1.0_2512.json
├── polyvore_outfit_v1.1_2512.json
├── mytheresa/
│   ├── images/
│   ├── looks-resized/
│   ├── ootd/
│   ├── edited/banana/
│   └── annotations/  # ATR, DensePose, DWPose, LIP, and mask-v3-look-resized
└── polyvore/         # Same asset layout

The v1.0 JSON index has 80,041 records. The recommended v1.1 index has 98,012 records: 97,068 train and 944 test, adding 17,971 training records. See the dataset card for counts by source, mask construction, and full asset details.

Prepare training data

Use the released v1.1 outfit annotations and retain their section assignments. Inpainting inputs replace the foreground of annotations/mask-v3-look-resized/<gender>/<id>.png with gray (128). The refined v3 masks improve coverage and include dilation; regenerate agnostic caches when changing masks. The dataset release includes ATR, DensePose, DWPose, LIP, and v3 mask annotations.

Editing source images are available at <subset>/edited/banana/<id>.png, and OOTD collages are included at <subset>/ootd/<gender>/<id>.png. Missing OOTD collages can be generated from reference item images in prompt order. The bundled example is a paired test sample for checking inference.

python scripts/data_gen/generate_2-refer.py --task inpainting --section train \
  --dataset-root /path/to/Garments2Look-data --num-samples 20000 \
  --num-workers 4 --seed 123 --output data/metadata/train-inpainting.json

# Requires editing source images.
python scripts/data_gen/generate_2-refer.py --task editing --section train \
  --dataset-root /path/to/Garments2Look-data --num-samples 20000 \
  --num-workers 4 --seed 123 --output data/metadata/train-editing.json

Metadata is a JSON list. Each record contains a target image, full prompt, and two edit_image paths in Figure 1 / Figure 2 order. Paths may be absolute or relative to DATASET_ROOT; generated metadata uses local paths, so regenerate it when moving data to another machine. Check the reported written/skipped counts before training.

[
  {
    "image": "path/to/target-look.png",
    "prompt": "Keep the person's identity, pose, background in Figure 1 unchanged, wearing the outfit in Figure 2...",
    "edit_image": ["path/to/source-person.png", "path/to/ootd.png"]
  }
]

LoRA training

Both Qwen-Image-Edit-2509 and Qwen-Image-2.1 use the same DiffSynth LoRA training workflow and two-reference metadata format. Garments2Look supplies dataset preparation, task inputs, and launch configurations; the model implementations and training algorithm come from DiffSynth. Both models use scripts/train/train.py, selected with --model_version. Inpainting and editing use separate metadata and produce separate adapters.

Install and extract the released runtime using the installation commands above. The training entry point explicitly imports models/Garments2Look-LoRA/qwen-runtime for both models. Shared settings are learning rate 1e-4, LoRA rank 32, BF16, gradient checkpointing, and a 1,048,576-pixel budget.

Configuration	Qwen-Image-Edit-2509	Qwen-Image-2.1
Pipeline	`QwenImagePipeline`	`QwenImage21Pipeline`
Target image processing	RGB; dimensions divisible by 16	RGBA; dimensions divisible by 32
Released training recipe	20,000 samples; 2 epochs	Full 97,068 training records; 1 epoch
Launcher	`train_lora.sh` with `TASK`	Task-specific `train_qwen21_lora-2-refer-*.sh`
Checkpoint defaults	End of each epoch	Every 1,000 steps and final step
Loss logs	`training.jsonl`	CSV logs

The LoRA target layers follow each model's architecture. For 2509: to_q,to_k,to_v,add_q_proj,add_k_proj,add_v_proj,to_out.0,to_add_out,img_mlp.net.2,img_mod.1,txt_mlp.net.2,txt_mod.1. For 2.1: to_q,to_k,to_v,to_out.0,img_mlp.proj,img_mlp.gate_layer,img_mlp.out,modulation.1. The 2.1 training configuration enables Flex Attention and disables KV caching. Both launchers enable find_unused_parameters.

Prepare metadata

Use the common data preparation instructions above for either model. The 20K commands reproduce the data count of the released 2509 recipe. For the full 2.1 split, use --num-samples 0:

export DATASET_ROOT=/path/to/Garments2Look-data
for task in inpainting editing; do
  python scripts/data_gen/generate_2-refer.py --task "$task" --section train \
    --dataset-root "$DATASET_ROOT" --num-samples 0 --num-workers 8 --seed 42 \
    --output "data/metadata/garments2look-2-refer-${task}-for-qwen21-train-97068.json"
done

Check that each full-split JSON contains 97,068 records with no skipped records. Metadata sampling seeds are separate from training and inference seeds.

Launch 2509

export MODEL_DIR="$PWD/models/Qwen-Image-Edit-2509"
export DATASET_ROOT=/path/to/Garments2Look-data
for task in inpainting editing; do
  TASK="$task" METADATA="$PWD/data/metadata/train-${task}.json" \
    OUTPUT_DIR="$PWD/models/train/qwen2509-${task}" \
    NPROC=1 SEED=123 bash scripts/train/train_lora.sh
done

Set NPROC=2 and CUDA_VISIBLE_DEVICES=0,1 for two GPUs. Override EPOCHS, LEARNING_RATE, NUM_WORKERS, or MASTER_PORT as needed. Epoch filenames are zero-indexed: epoch-1.safetensors is saved after the second epoch.

Launch 2.1

export MODEL_DIR=/path/to/Qwen-Image-2.1
export DATASET_ROOT=/path/to/Garments2Look-data
CUDA_VISIBLE_DEVICES=0,1 NPROC=2 bash scripts/train/train_qwen21_lora-2-refer-inpainting.sh
CUDA_VISIBLE_DEVICES=0,1 NPROC=2 bash scripts/train/train_qwen21_lora-2-refer-editing.sh

# Optional Slurm submission from the repository root, with the environment active.
inpainting_job=$(sbatch --parsable scripts/train/train_qwen21_lora-2-refer-inpainting.slurm)
sbatch --dependency=afterok:"$inpainting_job" scripts/train/train_qwen21_lora-2-refer-editing.slurm

The released 2.1 runs used two H200 GPUs with batch size 1 per process (global batch size 2), eight data workers, and no explicit training seed. An optional --seed can be passed to the launcher. Override METADATA_PATH, OUTPUT_PATH, NUM_EPOCHS, MAX_PIXELS, NUM_WORKERS, NPROC, or MASTER_PORT as needed. NPROC=1 supports a single GPU but changes the global batch size and step count. Adjust the Slurm templates for your cluster.

With 97,068 records and two processes, the final checkpoint is step-48534.safetensors; it is saved even though 48,534 is not a multiple of 1,000. The released HF adapters were renamed from this final checkpoint. Each completed H200 run took about 59 hours; timing depends on hardware.

License

The Garments2Look dataset is released under the Apache License 2.0, permitting commercial use, modification, and redistribution under its terms. Retain applicable copyright, attribution, and license notices, including those accompanying third-party materials.

The code is licensed under Apache 2.0 and includes an adapted DiffSynth-Studio implementation; see NOTICE. See the model repository for LoRA weight licensing information; an explicit adapter license is not yet specified.

Citation

If you use our dataset or models in your research, please consider citing our paper:

@inproceedings{cvpr2026garments2look,
    title={Garments2Look: A Multi-Reference Dataset for High-Fidelity Outfit-Level Virtual Try-On with Clothing and Accessories},
    author={Hu, Junyao and Cheng, Zhongwei and Wong, Waikeung and Zou, Xingxing},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2026}
}

Name		Name	Last commit message	Last commit date
diffsynth		diffsynth
docs		docs
examples		examples
scripts		scripts
.gitignore		.gitignore
LICENSE		LICENSE
NOTICE		NOTICE
README.md		README.md
dataset.py		dataset.py
pyproject.toml		pyproject.toml
requirements.txt		requirements.txt

Navigation Menu

Repository navigation

About

Repository files navigation

Garments2Look [CVPR 2026]

News and updates

TODO

Overview

Dataset construction

Qwen LoRA models

Checkpoints

Installation and download

Inpainting inference

Qwen-Image-Edit-2509

Qwen-Image-2.1

Editing inference

Qwen-Image-Edit-2509

Qwen-Image-2.1

Paired example: P00958796_b1

Qwen-Image-Edit-2509 LoRA

Qwen-Image-2.1 LoRA

Training details

Dataset preparation and training

Download training dataset

Prepare training data

LoRA training

Prepare metadata

Launch 2509

Launch 2.1

License

Citation

About

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Garments2Look [CVPR 2026]

News and updates

TODO

Overview

Dataset construction

Qwen LoRA models

Checkpoints

Installation and download

Inpainting inference

Qwen-Image-Edit-2509

Qwen-Image-2.1

Editing inference

Qwen-Image-Edit-2509

Qwen-Image-2.1

Paired example: P00958796_b1

Qwen-Image-Edit-2509 LoRA

Qwen-Image-2.1 LoRA

Training details

Dataset preparation and training

Download training dataset

Prepare training data

LoRA training

Prepare metadata

Launch 2509

Launch 2.1

License

Citation

About

Topics

Resources

Stars

Watchers

Forks

Contributors

Languages