lip-biting / README.md
abide1622's picture
Add the lip-biting training set (104 images) and dataset card
ebae14a verified
|
Raw History Blame Contribute Delete
1.78 kB
---
license: cc-by-nc-4.0
pretty_name: Lip Biting
task_categories:
- text-to-image
tags:
- lora
- lora-training
- image
- portrait
- facial-expression
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: data/train/*
---
# Lip Biting
104 portrait images for training a LoRA on a lip-biting expression.
Trigger word: **`lpb1te`**.
This is the training set exactly as it was used, published to get feedback
on curating it.
## Contents
| | |
| --- | --- |
| Images | 104 (102 PNG, 2 JPEG) |
| Typical size | 1024×1536 portrait (87 of the images). 12 are 1024×1024; the other 5 range from 3:2 landscape to 4:5 portrait. |
| Subjects | Women, head and shoulders, clothed, biting or pressing the lower lip, often smiling |
| Captions | None yet |
| Metadata | Stripped (EXIF removed from the two JPEGs; the PNGs carried none) |
## Usage
```python
from datasets import load_dataset
ds = load_dataset("abide1622/lip-biting", split="train") # one "image" column
```
To train, caption each image (a short description, plus `lpb1te` where the
expression should be learned) as same-named `.txt` files, or as a
`metadata.jsonl` with `file_name` and `text` columns in `data/train/`.
## Known limitations
- **No captions.** Every image teaches the trigger word the whole picture
(face, hair, lighting), not just the expression.
- **Narrow subject range.** All the subjects are young women, the framing is
similar throughout, and many share a warm studio look. A LoRA trained on
this set will carry that bias.
- **Near-duplicates.** Several images show the same subject and pose with
small variations.
## Feedback
Suggestions on what to remove, add or caption are welcome. Please open a discussion.