xcpan/VBench-I2V-832x480

VBench-I2V Images at 832x480

355 VBench-I2V input images prepared at 832 pixels wide by 480 pixels high, using the official annotation-guided 26:15 crop. Original image filenames and the official evaluation prompts are preserved.

This is a prepared derivative for convenient downloading, not an official VBench release. It contains input images and metadata; it does not contain generated evaluation videos.

Download

Download the complete image-and-metadata bundle without logging in:

curl -L --fail --retry 3 \
  https://huggingface.co/datasets/xcpan/VBench-I2V-832x480/resolve/main/vbench_i2v_832x480.zip \
  -o vbench_i2v_832x480.zip
curl -L --fail --retry 3 \
  https://huggingface.co/datasets/xcpan/VBench-I2V-832x480/resolve/main/vbench_i2v_832x480.zip.sha256 \
  -o vbench_i2v_832x480.zip.sha256
sha256sum -c vbench_i2v_832x480.zip.sha256
unzip vbench_i2v_832x480.zip -d vbench_i2v_832x480

Alternatively, download the individual files with the Hugging Face CLI:

hf download xcpan/VBench-I2V-832x480 --repo-type dataset \
  --include 'images/*' 'metadata/*' metadata.jsonl checksums.sha256 README.md \
  --local-dir vbench_i2v_832x480

The ZIP contains the images/ and metadata/ directories, metadata.jsonl, and their checksum manifest. After extracting it, verify the files with:

cd vbench_i2v_832x480
sha256sum -c checksums.sha256

Contents

Path	Contents
images/	All 355 RGB JPEGs, each exactly 832x480, with the original filenames
metadata/i2v-bench-info.json	Unmodified official image descriptions, source URLs, and crop annotations
metadata/vbench2_i2v_full_info.json	Unmodified official evaluation metadata: 1,118 prompts referencing the 355 images
metadata/target_crop_info_26-15.json	Derived crop boxes in original-image coordinates
metadata/provenance.json	Reproducibility information, source hashes, and preprocessing parameters
metadata/validation.json	Image-count, prompt-coverage, dimensions, and crop validation results
metadata.jsonl	Per-image records and source/output hashes; file_name paths are relative to the repository root
checksums.sha256	SHA256 checksums of the image and metadata payload
build_dataset.py	Script used to prepare and validate this dataset

Use images/<image_name> as the conditioning image for each entry in metadata/vbench2_i2v_full_info.json. The prompts are the original benchmark prompts, without prompt expansion. Five generated videos per image-prompt pair gives the full 5,590-video evaluation set.

Preprocessing

The upstream implementation is pinned to VBench commit fd18b3d055cb0fc6f066ca90fe2c3c8cbb698490:

1. Read the original image in its stored orientation; do not apply EXIF rotation before interpreting the annotations.
2. Call the official get_other_ratio_crop(second_crop, "26-15") and transfer_bbox_to_origin_img(first_crop, bbox) functions. The upstream algorithm seeds the random offset with 123 for each image.
3. Crop the original pixels using the resulting annotation-guided bounding box.
4. Resize to 832x480 with Pillow LANCZOS, then convert to RGB.
5. Save once as JPEG with quality=100 and subsampling=0, without carrying over EXIF orientation.

The crop geometry follows the official code exactly. The resize filter and final JPEG encoding are explicit choices for this prepared release; the upstream crop script itself does not specify this fixed output resolution.

To rebuild from the original images, install the versions in requirements.txt, obtain the original archive and extract its origin/ directory, then run:

git clone https://github.com/Vchitect/VBench.git VBench
git -C VBench checkout fd18b3d055cb0fc6f066ca90fe2c3c8cbb698490
python build_dataset.py \
  --origin-dir /path/to/origin \
  --crop-info metadata/i2v-bench-info.json \
  --upstream-root VBench \
  --output-dir rebuilt_vbench_i2v_832x480

The output directory must not already exist. Rebuilding the images uses only the CPU. Image hashes also depend on the Pillow/JPEG codec versions recorded in metadata/provenance.json; the public-source retrieval records in this release are additional provenance, not needed to compute the crop boxes.

Source provenance

• Official VBench-I2V documentation
• Official crop implementation at the pinned revision
• Official original-image archive on Google Drive
• Official image/crop annotations on Google Drive
• Original-image archive mirror used for transfer

The official ZIP download was quota-limited at preparation time. The original-image ZIP was transferred from the pinned public mirror above, not from an already resized image set. Its size is 632,512,065 bytes and its SHA256 is d8ce05f15a3313b3c61c9eda9f8515cd8e80f3b1f1799d3b0cd685a2c201f584. The image annotations were downloaded from the official source; the mirrored annotations were independently checked to be byte-identical. An original image downloaded directly from the official Drive folder was also checked against the corresponding archive member.

The official annotation SHA256 is 7dbe47684eadb2b4ace3fe452a4d8e8ceffa9c788d9c1ff048af3bb2585f96bf. The unmodified evaluation-metadata SHA256 is 8cf4d34dc11779a336c216a1a71d482927cfa24a1bf736ab5c92425609f129f8.

Licensing and attribution

VBench code is provided under its upstream Apache-2.0 license. This does not relicense the photographs. The images retain the terms of their original sources, including Pexels and Pixabay. Per-image source references are preserved in the original metadata. This derivative adds no broader rights to those images.

Some upstream source references are incomplete; they are preserved as supplied rather than replaced with guessed URLs. Consult the original source terms before further redistribution or reuse.

Please credit the original image sources and cite VBench and VBench++ when using this evaluation suite.