bat0man1/Twitter-X-Grok-Edit-Image-Pair-instruction
收藏资源简介:
--- dataset_info: features: - name: id dtype: string - name: control1_image dtype: image - name: control2_image dtype: image - name: target_image dtype: image - name: instruction dtype: string splits: - name: train num_bytes: 726215606 num_examples: 616 download_size: 726250417 dataset_size: 726215606 configs: - config_name: default data_files: - split: train path: data/train-* --- # Twitter-X-Grok-Edit-Image-Pair-instruction ## Dataset Summary **Twitter-X-Grok-Edit-Image-Pair-instruction** is a curated dataset of real-world image editing examples collected from Grok edits shared on X (Twitter). The dataset contains aligned edit instructions with **source images**, **reference images (when provided)**, and **final edited outputs**. It is designed for **instruction-guided image editing**, **multimodal alignment**, and **diffusion-based editing tasks**. The dataset will continue to expand with future updates. ## Dataset Structure Each sample contains: - **id** — filename / sample identifier - **control1_image** — base image (before edit) - **control2_image** — reference image used to guide the edit (present in applicable samples and should be used as provided) - **target_image** — edited (final) image - **instruction** — edit instruction describing the transformation ### Typical usage format (control1_image + instruction + control2_image when available) → target_image The dataset is distributed across **two `.parquet` files** inside the **data/** directory. Samples that include **control2_image** already contain it in their respective entries and should be used directly as part of the editing condition. ## Intended Use Suitable for: - instruction-based image editing models - diffusion editing pipelines - vision-language model training - multimodal alignment research ## Notes - Some samples include a **reference image (`control2_image`)** that is part of the edit condition and should be used as provided when present - Instructions are **manually written** - Images are **manually curated** from publicly shared Grok edits - The dataset will continue growing with future releases




