SimuScope: Realistic Endoscopic Synthetic Dataset Generation through Surgical Simulation and Diffusion Models

Contact Information

If you have any questions, suggestions, or would like to get in touch, feel free to contact me via email:

Email: s.martyniak@sanoscience.org

This is the official repository for SIMUSCOPE which has been accepted to IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2025, Tucson, AZ, USA. The dataset used in this project is publicly available on Zenodo. You can access and download it using the following link: https://zenodo.org/records/14205530

The workflow is divided into two steps:
1.Training
2.Inference

Training

The training was performed using https://github.com/bmaltais/kohya_ss.

To perform the training, follow the instructions provided at https://github.com/bmaltais/kohya_ss. In kohya_ss_CholectL45 and kohya_ss_CholectG_45, the necessary subfolders such as config, img, log, and model are already included. The img folder has already been generated. This setup is designed to be implemented in the kohya GUI.

Additionally, it is possible to skip using the kohya_ss GUI and directly upload an already generated command

Pretrained model it avaiable here to downoald: https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/blob/main/v1-5-pruned.safetensors

Training data

Data for trainig and pre-trained LoRAs you can find in folder Training (https://drive.google.com/drive/folders/186ItROA0xys9wkxQL9InKL9lJfyUY2Fp). You can also train LoRAs yourself:
-Training Lora CholectG45:

accelerate launch --num_cpu_threads_per_process=8 "/.../kohya_ss/sd-scripts/train_network.py" --network_train_unet_only --bucket_no_upscale --bucket_reso_steps=64 --cache_latents --cache_latents_to_disk --enable_bucket --min_bucket_reso=256 --max_bucket_reso=2048 --gradient_checkpointing --learning_rate="0.0012" --logging_dir="/.../WACV_2025/Training/Train_CholectG45/kohya_ss_CholectG45/log" --lr_scheduler="constant_with_warmup" --lr_scheduler_num_cycles="20" --max_data_loader_n_workers="0" --max_grad_norm="1" --resolution="512,512" --max_train_steps="13920" --min_snr_gamma=5 --mixed_precision="fp16" --network_alpha="1" --network_dim=128 --network_module=networks.lora --optimizer_type="Adafactor" --output_dir="/.../WACV_2025/Training/Train_CholectG45/kohya_ss_CholectG45/model" --output_name="Whole_s87_cholect45" --pretrained_model_name_or_path="/.../Pretrained_model_name/v1-5-pruned.safetensors" --save_every_n_epochs="1" --save_model_as=safetensors --save_precision="fp16" --text_encoder_lr=0.0012 --train_batch_size="5" --train_data_dir="/.../WACV_2025/Training/Train_CholectG45/kohya_ss_CholectG45/img" --unet_lr=0.0012 --xformers

-Training Lora CholectL45:

accelerate launch --num_cpu_threads_per_process=8 "/.../kohya_ss/sd-scripts/train_network.py" --network_train_unet_only --bucket_no_upscale --bucket_reso_steps=64 --cache_latents --cache_latents_to_disk --enable_bucket --min_bucket_reso=256 --max_bucket_reso=2048 --gradient_checkpointing --learning_rate="0.0012" --logging_dir="/.../WACV_2025/Training/Train_CholectL45/kohya_ss_CholectL45/log" --lr_scheduler="constant_with_warmup" --lr_scheduler_num_cycles="20" --max_data_loader_n_workers="0" --max_grad_norm="1" --resolution="512,512" --max_train_steps="13920" --min_snr_gamma=5 --mixed_precision="fp16" --network_alpha="1" --network_dim=128 --network_module=networks.lora --optimizer_type="Adafactor" --output_dir="/.../WACV_2025/Training/Train_CholectL45/kohya_ss_CholectL45/model" --output_name="Whole_s87_cholect45" --pretrained_model_name_or_path="/.../v1-5-pruned.safetensors" --save_every_n_epochs="1" --save_model_as=safetensors --save_precision="fp16" --text_encoder_lr=0.0012 --train_batch_size="5" --train_data_dir="/.../WACV_2025/Training/Train_CholectL45/kohya_ss_CholectL45/img" --unet_lr=0.0012 --xformers

Inference

We use webui for inference.To run inference we first have to run WebUI + ControlNet++ extension https://github.com/AUTOMATIC1111/stable-diffusion-webui

To use WebUI API: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/API

We use following ControlNet checkpoint:

• Reference ControlNet - We can use this after installing it in the WebUI: https://github.com/Mikubill/sd-webui-controlnet

To use ControlNet++ we have to download from: -SoftEdge and Depth https://huggingface.co/huchenlei/ControlNet_plus_plus_collection_fp16/tree/main

To run inference use api_inference.ipynb

Citation

This work was presented at the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2025. If you use this code or data, please cite our paper:

@InProceedings{Martyniak_2025_WACV,
    author    = {Martyniak, Sabina and Kaleta, Joanna and Dall'Alba, Diego and Naskr\k{e}t, Micha{\l} and P{\l}otka, Szymon and Korzeniowski, Przemys{\l}aw},
    title     = {SimuScope: Realistic Endoscopic Synthetic Dataset Generation through Surgical Simulation and Diffusion Models},
    booktitle = {Proceedings of the Winter Conference on Applications of Computer Vision (WACV)},
    month     = {February},
    year      = {2025},
    pages     = {4268-4278}
}