IsoVRBench

Paper · Code

Video preference pairs built by controlled intervention instead of human labeling. Each pair is a generated video and a counterpart that differs in one deliberate respect, so the construction fixes both the preferred side and the property being tested. Human validation checked once per intervention that the change is visible; no pair carries a human label.

split	pairs	clips	use
test	8,000	15,047	the IsoVRBench benchmark, 1,000 pairs per intervention
train	160,000	234,075, shared with validation	reward-model training, 20,000 pairs per intervention
validation	4,000		500 pairs per intervention

Splits are disjoint on prompt id, and every benchmark prompt is excluded from training and validation. Several interventions reuse one anchor clip, so a pair-level split would leak clips across splits.

Loading

from datasets import load_dataset
ds = load_dataset("stellalisy/IsoVRBench")   # train, validation, test

Each row names two clips. Test clips are files under videos/. Training and validation clips are packed into the tar shard named in each clip's shard field, stored under the same path.

import tarfile
from huggingface_hub import hf_hub_download

def clip_bytes(side, repo="stellalisy/IsoVRBench"):
    if side["shard"]:                                    # train / validation
        tar = hf_hub_download(repo, side["shard"], repo_type="dataset")
        with tarfile.open(tar) as tf:
            return tf.extractfile(side["path"]).read()
    path = hf_hub_download(repo, "videos/" + side["path"], repo_type="dataset")
    return open(path, "rb").read()                       # test

train/videos.jsonl maps every training and validation clip to its shard.

Record format

{
  "id": "iso-shuffle-0000",
  "split": "test",
  "intervention": "shuffle",
  "prompt": "three drawing characters running in a forest",
  "prompt_id": "p079459",
  "video_a": {"path": "gen2/p079459/s50.mp4", "prompt": "...", "shard": ""},
  "video_b": {"path": "gen2/p079459/shuffle.mp4", "prompt": "...", "shard": ""},
  "preferred": {"VQ": null, "MQ": "A", "TA": null},
  "scored_on": "MQ",
  "construction": "...",
  "details": ""
}

preferred gives the better side for each dimension the intervention changes, VQ visual quality, MQ motion quality and TA text alignment. A dimension the intervention leaves unchanged is null. The three generator substitutions change visual and motion quality together and carry both.

scored_on is set on test pairs only and names the dimension on which a model with separate heads is scored. details holds intervention parameters where they exist, such as the edited attribute of an ta_attrib pair, as a JSON string.

Test sides are randomized, so the preferred side is A in 50.6 percent of test pairs. Training sides are not randomized.

The eight interventions

The anchor is Wan2.2-T2V-A14B at 50 denoising steps, 832x480, 81 frames, 16 fps, guidance 5.0, seed 10.

intervention	paper name	what changes	dimension
shuffle	Temporal Shuffle	the anchor cut into six segments and reordered so that none stays in place	MQ
mq_hold	Frame Hold	8 to 12 anchor frames each held until the next, giving stuttering motion	MQ
mq_gen	Motion Re-gen	keyframes kept and the gaps regenerated by the anchor's generator	MQ
ta_attrib	Attribute Edit	one attribute of the prompt edited and rendered, then judged against the original prompt	TA
vq_steps	Reduced Steps	fewer denoising steps with the same model, prompt and seed, 8 in test and 16 in train and validation	VQ
size_i2v	Aligned Weak Gen	Wan2.2-TI2V-5B conditioned on the anchor's first frame	VQ, MQ
size_5b	Weak Gen (5B)	Wan2.2-TI2V-5B from the same prompt	VQ, MQ
size_1p3b	Weakest Gen (1.3B)	Wan2.1-T2V-1.3B from the same prompt	VQ, MQ

shuffle uses high-motion anchors, because reordering a nearly static clip changes nothing a viewer could see. The test split takes the 1,000 anchors with the most motion, and training and validation keep anchors above a fixed motion threshold.

Evaluating a reward model on the test split

Score every clip in videos.jsonl, one line per clip, then run the reference scorer. It reads isolated.jsonl, which holds the test split in its original format.

import json
with open("scores.jsonl", "w") as out:
    for line in open("videos.jsonl"):
        clip = json.loads(line)
        score = my_model(f"videos/{clip['path']}", clip["prompt"])
        out.write(json.dumps({"path": clip["path"], "prompt": clip["prompt"],
                              "score": float(score)}) + "\n")

python score.py --scores scores.jsonl --data . --single-head

A model with separate heads writes VQ, MQ and TA in place of score and drops --single-head, and each pair is then read on its scored_on head. Scores are keyed on (path, prompt).

score.py reports two conventions. ties_wrong counts a tie as wrong, since every pair carries a change people can see, and decided drops ties, as VideoGen-RewardBench does. The paper reports ties_wrong, averaged over the eight interventions.

Human validation

On 45 pairs per intervention, with same-clip controls interleaved, human raters agreed with the constructed label as below, counting a tie as wrong. Raters judged every control a tie.

size_5b	vq_steps	size_1p3b	size_i2v	ta_attrib	mq_gen	mq_hold	shuffle	average
98	84	100	98	98	100	93	100	96.4

Files

data/train.jsonl        160,000 training pairs
data/validation.jsonl   4,000 validation pairs
data/test.jsonl         8,000 test pairs
isolated.jsonl          the test pairs in the format score.py reads
videos.jsonl            the 15,047 test clips
videos/                 the test clips
train/videos.jsonl      the 234,075 training and validation clips and their shards
train/shards/           66 tar shards holding those clips
score.py                reference scorer for the test split
stats.json              counts by split and intervention

Provenance

Prompts come from VidProM. Every clip is generated by a model of the Wan2.1 or Wan2.2 family. Code for building and evaluating the pairs is at github.com/stellalisy/IsoVRBench.

Citation

@misc{li2026isovrbenchvideorewardmodeling,
      title={IsoVRBench: Video Reward Modeling Without Human Feedback}, 
      author={Shuyue Stella Li and Jacqueline He and Jiahui Chen and Sreya Dutta Roy and Yulia Tsvetkov and Luke Zettlemoyer and Marjan Ghazvininejad and Xiaochuang Han},
      year={2026},
      eprint={2610.17234},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.17234}, 
}

License

Prompts derive from VidProM and its terms carry through. The generated clips and all labels are released by the authors.