Datasets:
File size: 2,570 Bytes
b47c340 81b92d2 a41ba76 81b92d2 36b6dfa 81b92d2 a41ba76 74ab719 81b92d2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | ---
language:
- en
task_categories:
- text-to-image
- image-to-text
tags:
- image-captioning
---
# Meisho SFT
Meisho SFT is a dataset containing 10K curated and fully captioned, high quality images for the fine-tuning of diffusion models.
## Usage
```python
from datasets import load_dataset
dataset = load_dataset("lucidityai/meisho-sft", split="train")
```
## Curation
### Filtering
This dataset was filtered based via a two-stage process:
1. AI Detection
We used our Akita-1 AI image detection model to determine which images were AI generated with a 90% human made certainty to pass.
This cut a large portion of the data through what is most likely false positives, however, it also heavily reduces the risk of AI generated content appearing within the dataset.
2. VLM
We used Google's Gemma 4 26B model to determine if a image had any of the following qualities:
- Low quality (in the sense of poor editing)
- AI-likeness
- Fever-dreamy
- Watermarks
- NSFW-Content
### Annotations / Captioning
For creating image captions, we utilized Gemma 4 26B once more, this time giving it instructions to create a natural language prompt to recreate the target image 1:1, including detailing of any text to exacts.
For the data that included Danbooru tags, we included this into the prompt to provide supplementry information to the model so it could enrich it's prompt further.
## Safety and Risk
As with any synthetically curated dataset, the risk of incorrect captioning / annotating is worth noting in real world usage. Additional filtering may be neccesary if absolute quality is essensial for a training run / whatever else this may be used for.
This also means that watermarks, NSFW content and low quality content may appear in the dataset.
## Credits
The following datasets were used alongside some of our own collected data to create this dataset:
- [Animetimm's Danbooru-wdtagger-v4-w640-ws-150k](https://huggingface.co/datasets/animetimm/danbooru-wdtagger-v4-w640-ws-150k)
- [Krishnakamath's Pinterest_tag_prediction](https://huggingface.co/datasets/krishnakamath/pinterest_tag_prediction)
- [Kkcosmos's Instagram-images-with-captions](https://huggingface.co/datasets/kkcosmos/instagram-images-with-captions)
## Ethics and Opt-Out
Image generation is a highly sensitive area of AI/ML where the ability to be excluded matters, thus is why we release this dataset publically for individuals to check the list of images utilized. If you have an image in which you wish to have removed from this dataset, please contact imageteam@lucidity.sh. |