BZNR3/coco-pose-weather-edits

# COCO Pose Weather Edits

This upload package describes the local outputs for COCO pose-oriented weather editing generated with ControlNet.

## Local Source Data

- COCO subset: `/data2/baoming/coco2017/train1500_pose`
- Snow outputs: `/data2/baoming/controlnet/generated_coco_train1500_snow`
- Rain outputs: `/data2/baoming/controlnet/generated_coco_train1500_rain`

## Intended Hugging Face Layout

- `train1500_pose/`
- `generated_coco_train1500_snow/`
- `generated_coco_train1500_rain/`
- `README.md`

`train1500_pose/` contains:

- `images/`: the selected 1500-image subset
- `person_keypoints_train2017_1500.json`: filtered pose annotations
- `train1500_attrs.json`: image-level attribute mapping used to guide prompt construction

Each generated directory contains:

- `outputs/`: edited images
- `controls/`: ControlNet canny control maps
- `sample_1500_meta.json`: per-image prompt and generation metadata
- `logs/run.jsonl`: generation status log

## Counts

- `train1500_pose/images`: 1500 images
- `generated_coco_train1500_snow/outputs`: 1500 images
- `generated_coco_train1500_rain/outputs`: 1500 images

## Notes

- The subset is sampled from the first 20000 valid COCO `train2017` person-keypoint images covered by the BLIP/Gemini attribute file.
- Prompt construction preserves person count, placement, and body pose while editing weather conditions.