SanRiiiii/VitronEdit_8k

UniVEdit-8K

Paper | Project Page | Code

UniVEdit-8K is a dataset of multi-view editing pairs for 3D scenes. Each sample pairs a source multi-view sequence of a scene with an edited sequence of the same scene, together with the editing instruction.

The source sequences come from RealEstate10K clips (the 720p subset of RealCam-Vid). The editing instructions were generated with Qwen3-VL-30B-A3B, which describes the scene and selects the target object, and Qwen3.5-27B, which writes the instruction. The edited sequences were produced with Kling-O1.

Format

File	Content
<key>.source.mp4	source sequence: 72 frames, 24 fps, 1280×720
<key>.target.mp4	edited sequence: 72 frames, 24 fps, 1920×1080
<key>.txt	editing instruction, one English sentence

Loading

Both videos of a sample are returned as the bytes of the mp4 file. This helper decodes them into frames with PyAV:

import io
import av

def decode(mp4_bytes):
    with av.open(io.BytesIO(mp4_bytes)) as container:
        return [frame.to_ndarray(format="rgb24") for frame in container.decode(video=0)]

With Hugging Face datasets:

from datasets import load_dataset

ds = load_dataset("SanRiiiii/UniVEdit_8k", split="train", streaming=True)
for sample in ds:
    key = sample["__key__"]                 # "scene-..."
    instruction = sample["txt"]
    source = decode(sample["source.mp4"])   # 72 frames of shape (720, 1280, 3)
    target = decode(sample["target.mp4"])   # 72 frames of shape (1080, 1920, 3)

With webdataset:

import webdataset as wds

url = "https://huggingface.co/datasets/SanRiiiii/UniVEdit_8k/resolve/main/data/shard-{00000..00076}.tar"
for sample in wds.WebDataset(url, shardshuffle=False):
    key = sample["__key__"]
    instruction = sample["txt"].decode("utf-8")
    source = decode(sample["source.mp4"])
    target = decode(sample["target.mp4"])