Davide Morelli*, Alberto Baldrati*, Giuseppe Cartella, Marcella Cornia, Marco Bertini, Rita Cucchiara
* Equal contribution.
🔥🔥 [05/09/2023] Release of the training code
This is the official repository for the paper "LaDI-VTON: Latent Diffusion Textual-Inversion Enhanced Virtual Try-On".
Abstract:
The rapidly evolving fields of e-commerce and metaverse continue to seek innovative approaches to enhance the consumer experience. At the same time, recent advancements in the development of diffusion models have enabled generative networks to create remarkably realistic images. In this context, image-based virtual try-on, which consists in generating a novel image of a target model wearing a given in-shop garment, has yet to capitalize on the potential of these powerful generative solutions. This work introduces LaDI-VTON, the first Latent Diffusion textual Inversion-enhanced model for the Virtual Try-ON task. The proposed architecture relies on a latent diffusion model extended with a novel additional autoencoder module that exploits learnable skip connections to enhance the generation process preserving the model's characteristics. To effectively maintain the texture and details of the in-shop garment, we propose a textual inversion component that can map the visual features of the garment to the CLIP token embedding space and thus generate a set of pseudo-word token embeddings capable of conditioning the generation process. Experimental results on Dress Code and VITON-HD datasets demonstrate that our approach outperforms the competitors by a consistent margin, achieving a significant milestone for the task
If you make use of our work, please cite our paper:
@inproceedings{morelli2023ladi,
title={{LaDI-VTON: Latent Diffusion Textual-Inversion Enhanced Virtual Try-On}},
author={Morelli, Davide and Baldrati, Alberto and Cartella, Giuseppe and Cornia, Marcella and Bertini, Marco and Cucchiara, Rita},
booktitle={Proceedings of the ACM International Conference on Multimedia},
year={2023}
}We recommend using the Anaconda package manager to avoid dependency/reproducibility problems. For Linux systems, you can find a conda installation guide here.
- Clone the repository
git clone https://github.com/miccunifi/ladi-vton- Install Python dependencies
conda env create -n ladi-vton -f environment.yml
conda activate ladi-vtonAlternatively, you can create a new conda environment and install the required packages manually:
conda create -n ladi-vton -y python=3.10
conda activate ladi-vton
pip install torch==2.0.1 torchvision==0.15.2 opencv-python==4.7.0.72 diffusers==0.14.0 transformers==4.27.3 accelerate==0.18.0 clean-fid==0.1.35 torchmetrics[image]==0.11.4 wandb==0.14.0 matplotlib==3.7.1 tqdm xformers- Download the DressCode dataset
- To enhance the performance of our warping module, we have discovered that using in-shop images with a white background yields better results. To facilitate this process, we now offer pre-extracted masks that can be used to remove the background from the images. You can download the masks from the following link: here. Once downloaded, please extract the mask files and place them in the dataset folder alongside the corresponding images.
Once the dataset is downloaded, the folder structure should look like this:
├── DressCode
| ├── test_pairs_paired.txt
| ├── test_pairs_unpaired.txt
| ├── train_pairs.txt
│ ├── [dresses | lower_body | upper_body]
| | ├── test_pairs_paired.txt
| | ├── test_pairs_unpaired.txt
| | ├── train_pairs.txt
│ │ ├── images
│ │ │ ├── [013563_0.jpg | 013563_1.jpg | 013564_0.jpg | 013564_1.jpg | ...]
│ │ ├── masks
│ │ │ ├── [013563_1.png| 013564_1.png | ...]
│ │ ├── keypoints
│ │ │ ├── [013563_2.json | 013564_2.json | ...]
│ │ ├── label_maps
│ │ │ ├── [013563_4.png | 013564_4.png | ...]
│ │ ├── skeletons
│ │ │ ├── [013563_5.jpg | 013564_5.jpg | ...]
│ │ ├── dense
│ │ │ ├── [013563_5.png | 013563_5_uv.npz | 013564_5.png | 013564_5_uv.npz | ...]
- Download the VITON-HD dataset
Once the dataset is downloaded, the folder structure should look like this:
├── VITON-HD
| ├── test_pairs.txt
| ├── train_pairs.txt
│ ├── [train | test]
| | ├── image
│ │ │ ├── [000006_00.jpg | 000008_00.jpg | ...]
│ │ ├── cloth
│ │ │ ├── [000006_00.jpg | 000008_00.jpg | ...]
│ │ ├── cloth-mask
│ │ │ ├── [000006_00.jpg | 000008_00.jpg | ...]
│ │ ├── image-parse-v3
│ │ │ ├── [000006_00.png | 000008_00.png | ...]
│ │ ├── openpose_img
│ │ │ ├── [000006_00_rendered.png | 000008_00_rendered.png | ...]
│ │ ├── openpose_json
│ │ │ ├── [000006_00_keypoints.json | 000008_00_keypoints.json | ...]
To run the inference on the Dress Code or VITON-HD dataset, run the following command:
python src/inference.py --dataset [dresscode | vitonhd] --dresscode_dataroot <path> --vitonhd_dataroot <path> --output_dir <path> --test_order [paired | unpaired] --category [all | lower_body | upper_body | dresses ] --mixed_precision [no | fp16 | bf16] --enable_xformers_memory_efficient_attention --use_png --compute_metrics --dataset <str> dataset to use, options: ['dresscode', 'vitonhd']
--dresscode_dataroot <str> data root of dresscode dataset (required when dataset=dresscode)
--vitonhd_dataroot <str> data root of vitonhd dataset (required when dataset=vitonhd)
--test_order <str> test setting, options: ['paired', 'unpaired']
--category <str> category to test, options: ['all', 'lower_body', 'upper_body', 'dresses'] (default=all)
--output_dir <str> output directory
--batch_size <int> batch size (default=8)
--mixed_precision <str> mixed precision (no, fp16, bf16) (default=no)
--enable_xformers_memory_efficient_attention <store_true>
enable memory efficient attention in xformers (default=False)
--allow_tf32 <store_true> allow TF32 on Ampere GPUs (default=False)
--num_workers <int> number of workers (default=8)
--use_png <store_true> use png instead of jpg (default=False)
--compute_metrics compute metrics at the end of inference (default=False)
Since we release the pre-trained models via torch.hub, the models will be automatically downloaded when running the inference script.
Once you have run the inference script and extracted the images, you can compute the metrics by running the following command:
python src/utils/val_metrics.py --gen_folder <path> --dataset [dresscode | vitonhd] --dresscode_dataroot <path> --vitonhd_dataroot <path> --test_order [paired | unpaired] --category [all | lower_body | upper_body | dresses ] --gen_folder <str> Path to the generated images folder.
--dataset <str> dataset to use, options: ['dresscode', 'vitonhd']
--dresscode_dataroot <str> data root of dresscode dataset (required when dataset=dresscode)
--vitonhd_dataroot <str> data root of vitonhd dataset (required when dataset=vitonhd)
--test_order <str> test setting, options: ['paired', 'unpaired']
--category <str> category to test, options: ['all', 'lower_body', 'upper_body', 'dresses'] (default=all)
--batch_size batch size (default=32)
--workers number of workers (default=8)
In this section, you'll find instructions on how to train all the components of our model from scratch.
First of all, we need to train the warping module. To do so, run the following command:
python src/train_tps.py --dataset [dresscode | vitonhd] --dresscode_dataroot <path> --vitonhd_dataroot <path> --checkpoints_dir <path> --exp_name <str> --dataset <str> dataset to use, options: ['dresscode', 'vitonhd']
--dresscode_dataroot <str> dataroot of dresscode dataset (required when dataset=dresscode)
--vitonhd_dataroot <str> dataroot of vitonhd dataset (required when dataset=vitonhd)
--checkpoints_dir <str> checkpoints directory
--exp_name <str> experiment name
--batch_size <int> batch size (default=16)
--workers <int> number of workers (default=10)
--height <int> height of the input images (default=512)
--width <int> width of the input images (default=384)
--lr <float> learning rate (default=1e-4)
--const_weight <float> weight for the TPS constraint loss (default=0.01)
--wandb_log <store_true> log training on wandb (default=False)
--wandb_project <str> wandb project name (default=LaDI_VTON_tps)
--dense <store_true> use dense uv map instead of keypoints (default=False)
--only_extraction <store_true> only extract the images using the trained networks without training (default=False)
--vgg_weight <int> weight for the VGG loss (refinement network) (default=0.25)
--l1_weight <int> weight for the L1 loss (refinement network) (default=1.0)
--epochs_tps <int> number of epochs for the TPS training (default=50)
--epochs_refinement <int> number of epochs for the refinement network training (default=50)
At the end of the training, the warped cloth images will be saved in the data/warped_cloths
and data/warped_cloths_unpaired folders.
To save computation time, in the following steps, we will use the pre-extracted warped cloth images.
To train the EMASC module, run the following command:
python src/train_emasc.py --dataset [dresscode | vitonhd] --dresscode_dataroot <path> --vitonhd_dataroot <path> --output_dir <path> --dataset <str> dataset to use, options: ['dresscode', 'vitonhd']
--dresscode_dataroot <str> data root of dresscode dataset (required when dataset=dresscode)
--vitonhd_dataroot <str> data root of vitonhd dataset (required when dataset=vitonhd)
--output_dir <str> output directory where the model predictions and checkpoints will be written
--pretrained_model_name_or_path <str>
model identifier from huggingface.co/models (default=stabilityai/stable-diffusion-2-inpainting)
--seed <int> seed for reproducible training (default=1234)
--train_batch_size <int> batch size for training (default=16)
--test_batch_size <int> batch size for testing (default=16)
--num_train_epochs <int> number of training epochs (default=100)
--max_train_steps <int> maximum number of training steps. If provided, overrides num_train_epochs (default=40k)
--gradient_accumulation_steps <int>
number of update steps to accumulate before performing a backward/update pass (default=1)
--learning_rate <float> learning rate (default=1e-5)
--lr_scheduler <str> learning rate scheduler, options: ['linear', 'cosine', 'cosine_with_restarts', 'polynomial', 'constant', 'constant_with_warmup'] (default=constant_with_warmup)
--lr_warmup_steps <int> number of warmup steps for learning rate scheduler (default=500)
--allow_tf32 <store_true> allow TF32 on Ampere GPUs (default=False)
--adam_beta1 <float> value of beta_1 for Adam optimizer (default=0.9)
--adam_beta2 <float> value of beta_2 for Adam optimizer (default=0.999)
--adam_weight_decay <float> value of weight decay for Adam optimizer (default=1e-2)
--adam_epsilon <float> value of epsilon for Adam optimizer (default=1e-8)
--max_grad_norm <float> maximum value of gradient norm for gradient clipping (default=1.0)
--mixed_precision <str> mixed precision training, options: ['no', 'fp16', 'bf16'] (default=fp16)
--report_to <str> where to report metrics, options: ['wandb', 'tensorboard', 'comet_ml'] (default=wandb)
--checkpointing_steps <int> number of steps between each checkpoint (default=10000)
--resume_from_checkpoint <str> whether training should be resumed from a previous checkpoint. Use a "latest" to automatically select the last available checkpoint. (default=None)
--num_workers <int> number of workers (default=8)
--num_workers_test <int> number of workers for test dataloader (default=8)
--test_order <str> test setting, options: ['paired', 'unpaired'] (default=paired)
--emasc_type <str> type of EMASC, options: ['linear', 'nonlinear'] (default=nonlinear)
--vgg_weight <float> weight for the VGG loss (default=0.5)
--emasc_kernel <int> kernel size for the EMASC module (default=3)
--emasc_padding <int> padding for the EMASC module (default=1)
At the end of the training, the EMASC checkpoints will be saved in the output_dir folder.
To accelerate the training process for subsequent steps, consider pre-computing the CLIP cloth embeddings for each image in the dataset.
To do so, run the following command:
python src/utils/compute_cloth_clip_features.py --dataset [dresscode | vitonhd] --dresscode_dataroot <path> --vitonhd_dataroot <path> --dataset <str> dataset to use, options: ['dresscode', 'vitonhd']
--dresscode_dataroot <str> data root of dresscode dataset (required when dataset=dresscode)
--vitonhd_dataroot <str> data root of vitonhd dataset (required when dataset=vitonhd)
--pretrained_model_name_or_path <str>
model identifier from huggingface.co/models (default=stabilityai/stable-diffusion-2-inpainting)
--batch_size <int> batch size (default=16)
--num_workers <int> number of workers (default=8)
The computed features will be saved in the data/clip_cloth_embeddings folder.
In the following steps, to use the pre-computed features, make sure to use the --use_clip_cloth_features flag.
To pre-train the inversion adapter, run the following command:
python src/train_inversion_adapter.py --dataset [dresscode | vitonhd] --dresscode_dataroot <path> --vitonhd_dataroot <path> --output_dir <path> --gradient_checkpointing --enable_xformers_memory_efficient_attention --use_clip_cloth_features --dataset <str> dataset to use, options: ['dresscode', 'vitonhd']
--dresscode_dataroot <str> data root of dresscode dataset (required when dataset=dresscode)
--vitonhd_dataroot <str> data root of vitonhd dataset (required when dataset=vitonhd)
--output_dir <str> output directory where the model predictions and checkpoints will be written
--pretrained_model_name_or_path <str>
model identifier from huggingface.co/models (default=stabilityai/stable-diffusion-2-inpainting)
--seed <int> seed for reproducible training (default=1234)
--train_batch_size <int> batch size for training (default=16)
--test_batch_size <int> batch size for testing (default=16)
--num_train_epochs <int> number of training epochs (default=100)
--max_train_steps <int> maximum number of training steps. If provided, overrides num_train_epochs (default=200k)
--gradient_accumulation_steps <int>
number of update steps to accumulate before performing a backward/update pass (default=1)
--gradient_checkpointing <store_true>
use gradient checkpointing to save memory at the expense of slower backward pass (default=False)
--learning_rate <float> learning rate (default=1e-5)
--lr_scheduler <str> learning rate scheduler, options: ['linear', 'cosine', 'cosine_with_restarts', 'polynomial', 'constant', 'constant_with_warmup'] (default=constant_with_warmup)
--lr_warmup_steps <int> number of warmup steps for learning rate scheduler (default=500)
--allow_tf32 <store_true> allow TF32 on Ampere GPUs (default=False)
--adam_beta1 <float> value of beta_1 for Adam optimizer (default=0.9)
--adam_beta2 <float> value of beta_2 for Adam optimizer (default=0.999)
--adam_weight_decay <float> value of weight decay for Adam optimizer (default=1e-2)
--adam_epsilon <float> value of epsilon for Adam optimizer (default=1e-8)
--max_grad_norm <float> maximum value of gradient norm for gradient clipping (default=1.0)
--mixed_precision <str> mixed precision training, options: ['no', 'fp16', 'bf16'] (default=fp16)
--report_to <str> where to report metrics, options: ['wandb', 'tensorboard', 'comet_ml'] (default=wandb)
--checkpointing_steps <int> number of steps between each checkpoint (default=50000)
--resume_from_checkpoint <str> whether training should be resumed from a previous checkpoint. Use a "latest" to automatically select the last available checkpoint. (default=None)
--enable_xformers_memory_efficient_attention <store_true>
enable memory efficient attention in xformers (default=False)
--num_workers <int> number of workers (default=8)
--num_workers_test <int> number of workers for test dataloader (default=8)
--test_order <str> test setting, options: ['paired', 'unpaired'] (default=paired)
--num_vstar <int> number of predicted v* per image to use (default=16)
--num_encoder_layers <int> number of ViT layers to use in inversion adapter (default=1)
--use_clip_cloth_features <store_true>
use precomputed clip cloth features instead of computing them each iteration (default=False).
At the end of the training, the inversion adapter checkpoints will be saved in the output_dir folder.
NOTE: You can use the --use_clip_cloth_features flag only if you have previously computed the clip cloth features
using the src/utils/compute_cloth_clip_features.py script (step 2.5).
To successfully train the VTO model, ensure that you specify the correct path to the pre-trained inversion adapter
checkpoint. If omitted, the inversion adapter will be trained from scratch. Additionally, don't forget to include the
--train_inversion_adapter flag to enable the inversion adapter training during the VTO training process.
To train the VTO model, run the following command:
python src/train_vto.py --dataset [dresscode | vitonhd] --dresscode_dataroot <path> --vitonhd_dataroot <path> --output_dir <path> --inversion_adapter_dir <path> --gradient_checkpointing --enable_xformers_memory_efficient_attention --use_clip_cloth_features --train_inversion_adapter --dataset <str> dataset to use, options: ['dresscode', 'vitonhd']
--dresscode_dataroot <str> data root of dresscode dataset (required when dataset=dresscode)
--vitonhd_dataroot <str> data root of vitonhd dataset (required when dataset=vitonhd)
--output_dir <str> output directory where the model predictions and checkpoints will be written
--inversion_adapter_dir <str> path to the inversion adapter checkpoint directory. Should be the same as `output_dir` of the inversion adapter training script. If not specified, the inversion adapter will be trained from scratch. (default=None)
--inversion_adapter_name <str> name of the inversion adapter checkpoint. To load the latest checkpoint, use `latest`. (default=latest)
--pretrained_model_name_or_path <str>
model identifier from huggingface.co/models (default=stabilityai/stable-diffusion-2-inpainting)
--seed <int> seed for reproducible training (default=1234)
--train_batch_size <int> batch size for training (default=16)
--test_batch_size <int> batch size for testing (default=16)
--num_train_epochs <int> number of training epochs (default=100)
--max_train_steps <int> maximum number of training steps. If provided, overrides num_train_epochs (default=200k)
--gradient_accumulation_steps <int>
number of update steps to accumulate before performing a backward/update pass (default=1)
--gradient_checkpointing <store_true>
use gradient checkpointing to save memory at the expense of slower backward pass (default=False)
--learning_rate <float> learning rate (default=1e-5)
--lr_scheduler <str> learning rate scheduler, options: ['linear', 'cosine', 'cosine_with_restarts', 'polynomial', 'constant', 'constant_with_warmup'] (default=constant_with_warmup)
--lr_warmup_steps <int> number of warmup steps for learning rate scheduler (default=500)
--allow_tf32 <store_true> allow TF32 on Ampere GPUs (default=False)
--adam_beta1 <float> value of beta_1 for Adam optimizer (default=0.9)
--adam_beta2 <float> value of beta_2 for Adam optimizer (default=0.999)
--adam_weight_decay <float> value of weight decay for Adam optimizer (default=1e-2)
--adam_epsilon <float> value of epsilon for Adam optimizer (default=1e-8)
--max_grad_norm <float> maximum value of gradient norm for gradient clipping (default=1.0)
--mixed_precision <str> mixed precision training, options: ['no', 'fp16', 'bf16'] (default=fp16)
--report_to <str> where to report metrics, options: ['wandb', 'tensorboard', 'comet_ml'] (default=wandb)
--checkpointing_steps <int> number of steps between each checkpoint (default=50000)
--resume_from_checkpoint <str> whether training should be resumed from a previous checkpoint. Use a "latest" to automatically select the last available checkpoint. (default=None)
--enable_xformers_memory_efficient_attention <store_true>
enable memory efficient attention in xformers (default=False)
--num_workers <int> number of workers (default=8)
--num_workers_test <int> number of workers for test dataloader (default=8)
--test_order <str> test setting, options: ['paired', 'unpaired'] (default=paired)
--uncond_fraction <float> fraction of unconditioned training samples (default=0.2)
--text_usage <str> text features to use, options: ['none', 'noun_chunks', 'inversion_adapter'] (default=inversion_adapter)
--cloth_input_type <str> cloth input type, options: ['none', 'warped'], (default=warped)
--num_vstar <int> number of predicted v* per image to use (default=16)
--num_encoder_layers <int> number of ViT layers to use in inversion adapter (default=1)
--train_inversion_adapter <store_true>
train the inversion adapter during the VTO training (default=False)
--use_clip_cloth_features <store_true>
use precomputed clip cloth features instead of computing them each iteration (default=False).
At the end of the training, the checkpoints will be saved in the output_dir folder.
NOTE: You can use the --use_clip_cloth_features flag only if you have previously computed the clip cloth features
using the src/utils/compute_cloth_clip_features.py script (step 2.5).
Before running the inference, make sure to specify the correct path to all the trained checkpoints. Make sure to also use coherent hyperparameters with the ones used during training.
To run the inference on the Dress Code or VITON-HD dataset, run the following command:
python src/eval.py --dataset [dresscode | vitonhd] --dresscode_dataroot <path> --vitonhd_dataroot <path> --output_dir <path> --save_name <str> --test_order [paired | unpaired] --unet_dir <path> --inversion_adapter_dir <path> --emasc_dir <path> --category [all | lower_body | upper_body | dresses ] --enable_xformers_memory_efficient_attention --use_png --compute_metrics --dataset <str> dataset to use, options: ['dresscode', 'vitonhd']
--dresscode_dataroot <str> data root of dresscode dataset (required when dataset=dresscode)
--vitonhd_dataroot <str> data root of vitonhd dataset (required when dataset=vitonhd)
--output_dir <str> output directory where the generated images will be written
--save_name <str> name of the generated images folder inside `output_dir`
--test_order <str> test setting, options: ['paired', 'unpaired']
--unet_dir <str> path to the UNet checkpoint directory. Should be the same as `output_dir` of the VTO training script
--unet_name <str> name of the UNet checkpoint. To load the latest checkpoint, use `latest`. (default=latest)
--inversion_adapter_dir <str> path to the inversion adapter checkpoint directory. Should be the same as `output_dir` of the VTO training script. Needed only if `--text_usage` is set to `inversion_adapter`. (default=None)
--inversion_adapter_name <str> name of the inversion adapter checkpoint. To load the latest checkpoint, use `latest`. (default=latest)
--emasc_dir <str> path to the EMASC checkpoint directory. Should be the same as `output_dir` of the EMASC training script. Needed when --emasc_type!=none. (default=None)
--emasc_name <str> name of the EMASC checkpoint. To load the latest checkpoint, use `latest`. (default=latest)
--pretrained_model_name_or_path <str>
model identifier from huggingface.co/models (default=stabilityai/stable-diffusion-2-inpainting)
--seed <int> seed for reproducible training (default=1234)
--batch_size <int> batch size(default=8)
--allow_tf32 <store_true> allow TF32 on Ampere GPUs (default=False)
--enable_xformers_memory_efficient_attention <store_true>
enable memory efficient attention in xformers (default=False)
--num_workers <int> number of workers (default=8)
--category <str> category to test, options: ['all', 'lower_body', 'upper_body', 'dresses'] (default=all)
--emasc_type <str> type of EMASC, options: ['linear', 'nonlinear'] (default=nonlinear)
--emasc_kernel <int> kernel size for the EMASC module (default=3)
--emasc_padding <int> padding for the EMASC module (default=1)
--text_usage <str> text features to use, options: ['none', 'noun_chunks', 'inversion_adapter'] (default=inversion_adapter)
--cloth_input_type <str> cloth input type, options: ['none', 'warped'], (default=warped)
--num_vstar <int> number of predicted v* per image to use (default=16)
--num_encoder_layers <int> number of ViT layers to use in inversion adapter (default=1)
--use_png <store_true> use png instead of jpg (default=False)
--num_inference_steps <int> number of diffusion steps at inference time (default=50)
--guidance_scale <float> guidance scale of the diffusion (default=7.5)
--use_clip_cloth_features <store_true>
use precomputed clip cloth features instead of computing them each iteration (default=False).
--compute_metrics compute metrics at the end of inference (default=False)
The generated images will be saved in the output_dir/save_name_{test_order} folder.
NOTE: You can use the --use_clip_cloth_features flag only if you have previously computed the clip cloth features
using the src/utils/compute_cloth_clip_features.py script (step 2.5).
This work has partially been supported by the PNRR project “Future Artificial Intelligence Research (FAIR)”, by the PRIN project “CREATIVE: CRoss-modal understanding and gEnerATIon of Visual and tExtual content” (CUP B87G22000460001), both co-funded by the Italian Ministry of University and Research, and by the European Commission under European Horizon 2020 Programme, grant number 101004545 - ReInHerit.
All material is made available under Creative Commons BY-NC 4.0. You
can use, redistribute, and adapt the material for non-commercial purposes, as long as you give appropriate
credit by citing our paper and indicate any changes that you've made.
