ego2act/ego2act-vidgen

Ego2Act Generated Videos and Scores

Paper · Project page · Code · Benchmark tasks and human recordings

Videos generated by six image-to-video models on the Ego2Act benchmark, three seeds each, together with every score reported in the paper, namely Ego2ActJudge, human ratings and five baseline evaluators plus a single-call VQA control. All tables share one key, video_id.

Model	Access
Seedance 2.0 (ByteDance)	OpenRouter
Kling 3.0 Pro (Kuaishou)	OpenRouter
Wan 2.7 (Alibaba)	OpenRouter
Grok Imagine 1.5 (xAI)	OpenRouter
MiniMax H3 (MiniMax)	open weights, local server
Cosmos 3 Nano (NVIDIA)	open weights, local server, the Nano variant of Cosmos 3

Quick start

from datasets import load_dataset

judge = load_dataset("ego2act/ego2act-vidgen", "judge_scores", split="train").to_pandas()
human = load_dataset("ego2act/ego2act-vidgen", "human_ratings", split="train").to_pandas()
both = human.merge(judge, on="video_id", suffixes=("_human", "_judge"))
print(both[["final_human", "final_judge"]].corr())

# one model only
huggingface-cli download ego2act/ego2act-vidgen --repo-type dataset --include "videos/wan_2_7/*" --local-dir ego2act

Subsets

videos (default), one row per generated video

Column	Type	Description
file	string	path of the mp4 inside this repo (480p H.264, 24 fps, no audio)
url	string	direct download link to the mp4
video_id	string	<case_id>::<model>__<seed>
case_id	string	task in ego2act-bench
model	string	generator
seed	int	101, 202 or 303
duration_s, width, height	float, int	measured from the file

judge_scores, Ego2ActJudge on every video

Column	Description
video_id, case_id, model, seed	video; model and seed are empty for human recordings
source	generated, human_correct or human_wrong
task, physics, final	0–100; Final = √(Task · Physics); empty when the judge could not score the video

Human recordings are scored too, and live in ego2act-bench.

human_ratings, the 600-video human panel

Column	Description
video_id, case_id, source, model	video (25 tasks, generated videos plus human correct and wrong recordings)
rater_{1,2,3}_task	each rater's Task level, 0–3
rater_{1,2,3}_physics	each rater's Physics level, 0–4
task, physics, final	aggregate on 0–100
rater_count	number of raters for the video

Raters are anonymised.

baseline_scores, other evaluators on the same videos

Column	Description
video_id, case_id	video
evaluator	wr-arena, pqsg, rbench, worldmodelbench, videoscore or simple_vqa
task, physics	evaluator output mapped to 0–100 (WR-Arena scores Task only)
final	evaluator's own overall score mapped to 0–100
native_raw	the evaluator's raw output as JSON

traces/, the evidence behind the scores

Plain files that join to the tables on video_id (and evaluator for the baselines).

Path	Content
traces/ego2act_judge/traces.jsonl	Ego2ActJudge gate-by-gate evidence for every scored video
traces/ego2act_judge/plans.jsonl	the judge's Task and Physics plans, keyed by video_id, axis and run
traces/baselines/<evaluator>/	scores.csv and traces.jsonl of each baseline evaluator on the human-rated videos

How scores are produced

Ego2ActJudge (Gemini 3.7 Flash through OpenRouter) splits the goal into subgoals and checks each with Task gates T1–T3 and Physics gates P1–P4, stopping at the first failure. It needs no reference video. Human raters used the same rubric. Baselines run on their official code at pinned commits, see the code repository.

Intended use and limitations

Use these tables to compare video generators on goal-directed manipulation and to measure how well an automatic judge agrees with people. Scores are for the model versions and settings listed in the code repository and can drift as hosted models change. The judge is more lenient on Physics than human raters, by about 24 points on average.

License

CC BY 4.0. Generated videos are released for research on evaluation. The code is MIT licensed.

Citation

@article{ego2act2026,
  title   = {Ego2Act: Evaluating Goal-Directed Manipulation in Egocentric Video Generation},
  author  = {Irawan, Patrick Amadeus and Parlambang, Iskandar Muda Rizky and Maulana, Rava and Cui, Qinrong and Fuadi, Erland Hilman and Zuhri, Zayd M. K. and Absar, Nanda Ryaas and Elshabrawy, Ahmed and Mulyawan, Wilfried Ariel and Yu, Shoubin and Zhang, Yue and Bansal, Mohit and Aji, Alham Fikri},
  journal = {arXiv preprint arXiv:2610.01092},
  year    = {2026}
}