Garments2Look-LoRA

Task-specific DiffSynth LoRA adapters for Qwen-Image-Edit-2509 and Qwen-Image-2.1, trained for outfit-level virtual try-on with clothing and accessories.

• Code and dataset preparation
• Dataset
• Paper (CVPR 2026)

This model repository includes the four adapters, a compatible Qwen inference runtime, one unified inference script, and a paired test example. The runtime is a trimmed DiffSynth 2.1.8 implementation used for these results. The script below supports both base models; no GitHub update is required to run it.

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 manifest.json.

Inputs

Both tasks use the same two-reference interface:

Input	Inpainting	Editing
Figure 1	Person with clothing regions already masked gray (128)	Source person wearing an existing outfit
Figure 2	OOTD collage of target products	OOTD collage of target products
Prompt	Numbered items, styling, and layering order	Numbered items, styling, and layering order

No separate mask argument is passed to inference. Match the numbered items to the collage. For dataset-based inpainting, use the updated v3 masks with dilation. The dataset has 98,012 total records: 97,068 training and 944 test records, with v1.1 annotations and editing source images. See the dataset card for preparation.

Installation and download

Tested with Python 3.10, PyTorch 2.7.1 / CUDA 12.8, Transformers 5.16.1, and H200 GPUs. Run all commands below from the same working directory. Download only the base model you intend to use.

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 bundled script imports the extracted runtime. Adapters use the DiffSynth format; compatibility with other LoRA loaders has not been verified. The Hub diffusers library tag enables download tracking for the root-level safetensors files; it does not indicate verified Diffusers loader compatibility. Set CUDA_VISIBLE_DEVICES to choose a GPU.

Inpainting inference

Qwen-Image-Edit-2509

python models/Garments2Look-LoRA/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 models/Garments2Look-LoRA/examples/qwen2509/input-inpainting.png \
  --ootd models/Garments2Look-LoRA/examples/qwen2509/ootd.png \
  --prompt-file models/Garments2Look-LoRA/examples/qwen2509/prompt-inpainting.txt \
  --output output/qwen2509-inpainting.png --seed 0 --steps 40 --cfg-scale 4

Qwen-Image-2.1

python models/Garments2Look-LoRA/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 models/Garments2Look-LoRA/examples/qwen21/input-inpainting.png \
  --ootd models/Garments2Look-LoRA/examples/qwen21/ootd.png \
  --prompt-file models/Garments2Look-LoRA/examples/qwen21/prompt-inpainting.txt \
  --output output/qwen21-inpainting.png --seed 0 --steps 40 --cfg-scale 1

Editing inference

Qwen-Image-Edit-2509

python models/Garments2Look-LoRA/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 models/Garments2Look-LoRA/examples/qwen2509/input-editing.png \
  --ootd models/Garments2Look-LoRA/examples/qwen2509/ootd.png \
  --prompt-file models/Garments2Look-LoRA/examples/qwen2509/prompt-editing.txt \
  --output output/qwen2509-editing.png --seed 0 --steps 40 --cfg-scale 4

Qwen-Image-2.1

python models/Garments2Look-LoRA/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 models/Garments2Look-LoRA/examples/qwen21/input-editing.png \
  --ootd models/Garments2Look-LoRA/examples/qwen21/ootd.png \
  --prompt-file models/Garments2Look-LoRA/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 bundled runtime and script here reproduce inference for both models; this package does not include a complete 2.1 training launcher.

Repository layout

Garments2Look-LoRA/
├── README.md
├── manifest.json
├── inference.py
├── qwen-runtime.zip
├── examples/
│   ├── selection.json
│   ├── qwen2509/  # inputs, prompts, target, outputs, parameters, comparison
│   └── qwen21/    # same outfit; final 2.1 outputs
├── Qwen-Image-Edit-2509-LoRA-2-refer-20k-*-epoch-1.safetensors
└── Qwen-Image-2.1-LoRA-2-refer-97068-*-epoch-0.safetensors

Limitations and license

Fine accessory details, garment fidelity, styling, layering, and pose preservation vary with input and seed. The showcased example is selected, not an aggregate evaluation. Full base models are required, and their respective licenses apply separately. The bundled DiffSynth code retains its upstream license; an explicit adapter license has not yet been 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}
}