metadata
license: mit
dataset_info:
features:
- name: id
dtype: string
- name: control_images
list: image
- name: control_mask
dtype: image
- name: target_image
dtype: image
- name: prompt
dtype: string
splits:
- name: train
num_bytes: 8458819
num_examples: 20
- name: test
num_bytes: 2807087
num_examples: 2
download_size: 11241583
dataset_size: 11265906
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
task_categories:
- image-to-image
language:
- en
tags:
- face_seg
pretty_name: tsien
size_categories:
- n<1K
This is the dataset used for FluxKontext or Qwen-Image-Edit training with the task image-text2image. This is a good adtaset that can very fine the effect of finetuning, where the base model is not good at
Refer this for the lora finetune model qwen-image-edit-lora-face-segmentation
Usage of the dataset
dd = load_editing_dataset("TsienDragon/face_segmentation_20")
sample = dd["test"][0]
Example of the data structure
DATASET_ROOT/train
├── control_images
│ ├── 060002_4_028450_FEMALE_30.png
│ ├── 060002_4_028450_FEMALE_30_1.png
│ ├── 060002_4_028450_FEMALE_30_2.jpg
│ ├── 060002_4_028450_FEMALE_30_mask.png
│ ├── 060003_4_028451_FEMALE_65.png
│ ├── 060003_4_028451_FEMALE_65_1.png
│ └── 060003_4_028451_FEMALE_65_mask.png
└── training_images
├── 060002_4_028450_FEMALE_30.txt
├── 060002_4_028450_FEMALE_30.png
├── 060003_4_028451_FEMALE_65.txt
└── 060003_4_028451_FEMALE_65.png