Visual-Intelligence / README.md
reiofa's picture
Update README.md
c015ed1 verified
---
dataset_info:
features:
- name: id
dtype: int64
- name: input
list:
- name: type
dtype: string
- name: content
dtype: string
- name: output
struct:
- name: veo3
list: string
- name: framepack
list: string
- name: framepack_seleted_video
dtype: string
- name: hunyuan
list: string
- name: hunyuan_seleted_video
dtype: string
- name: wan2.2-14b
list: string
- name: wan2.2-14b_seleted_video
dtype: string
- name: wan2.2-5b
list: string
- name: wan2.2-5b_seleted_video
dtype: string
splits:
- name: train
num_bytes: 98746
num_examples: 99
download_size: 36034
dataset_size: 98746
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---
# Visual-Intelligence
## πŸ”— Links
- [πŸ’Ύ Github Repo](https://github.com/Entroplay/Visual-Intelligence)
- [πŸ€— HF Dataset](https://huggingface.co/datasets/Entroplay/Visual-Intelligence)
- [πŸ“‘ Blog](https://entroplay.ai/research/video-intelligence)
## πŸ“– Dataset Introduction
### Dataset Schema
- **id**: Unique sample identifier.
- **input**: Ordered list describing the input context.
- **type**: Either "image" or "text".
- **content**: For "image", a relative path to the first-frame image. For "text", the prompt text.
- **output**: Generated candidates and final selections by model.
- **veo3**: Relative paths to videos generated by the VEO3 pipeline.
- **framepack**: Relative paths to videos generated by FramePack across multiple runs.
- **hunyuan**: Relative paths to videos generated by Hunyuan across multiple runs.
- **wan2.2-5b**: Relative paths to videos generated by Wan-2.2-5B across multiple runs.
- **wan2.2-14b**: Relative paths to videos generated by Wan-2.2-14B across multiple runs.
- **framepack_seleted_video**: Selected best video among FramePack candidates.
- **hunyuan_seleted_video**: Selected best video among Hunyuan candidates.
- **wan2.2-5b_seleted_video**: Selected best video among Wan 2.2 5B candidates.
- **wan2.2-14b_seleted_video**: Selected best video among Wan 2.2 14B candidates.
### Data Format:
```json
{
"id": 1,
"input": [
{ "type": "image", "content": "thumbnails/mp4/keypoint_localization.jpg" },
{ "type": "text", "content": "Add a bright blue dot at the tip of the branch on which the macaw is sitting. ..." }
],
"output": {
"veo3": ["videos/mp4/keypoint_localization.mp4"],
"framepack": [
"videos/1_framepack_1.mp4",
"videos/1_framepack_2.mp4"
],
"hunyuan": [
"videos/1_hunyuan_1.mp4",
"videos/1_hunyuan_2.mp4"
],
"wan2.2-5b": [
"videos/1_wan2.2-5b_1.mp4",
"videos/1_wan2.2-5b_2.mp4"
],
"wan2.2-14b": [
"videos/1_wan2.2-14b_1.mp4",
"videos/1_wan2.2-14b_2.mp4"
],
"framepack_seleted_video": "videos/1_framepack_1.mp4",
"hunyuan_seleted_video": "videos/1_hunyuan_1.mp4",
"wan2.2-5b_seleted_video": "videos/1_wan2.2-5b_1.mp4",
"wan2.2-14b_seleted_video": "videos/1_wan2.2-14b_1.mp4"
}
}
```
## πŸš€ About project
Google' Veo 3 shows extreme promise in visual intelligence, demonstrating strong visual commonsense and reasoning in visual generation. We aim to construct a fully open-source evaluation suite to measure current progress in video generative intelligence across various dimensions among several state-of-the-art proprietary and open-source models.