Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
video
video
5
5

Physics-IQ Verified V2V — 물리 범주별 분석 (Cosmos3 Nano, Cosmos-Predict2.5-2B)

Physics-IQ Verified V2V 점수가 어떤 물리 범주(Rigid / Non-rigid Solid / Fluid / Optics / Thermodynamics / Magnetism)에서 나오는지 보려고, 두 모델을 공식 채점기로 채점하고 범주별로 나눴다.

  • Cosmos3 Nano: 공개된 생성 영상(Astra 프롬프트 업샘플링, 4 runs)을 받아 다시 채점했다. 영상은 이 저장소에 없다(§4의 원본 링크 참고).
  • Cosmos-Predict2.5-2B: 198개 V2V 영상을 직접 생성해서 채점했다. 영상은 cosmos-predict2.5-2b/videos/에 있다.

작업일: 2026-11-08 ~ 11-09 (KST). 서버: RTX 3090 × 8 (driver 525.147.05).


1. 결과 요약 (Verified V2V, 0–100)

범주 (장면 / 샘플) Cosmos3 Nano (Astra, 4 runs) Cosmos-Predict2.5-2B (1 run)
Overall (66 / 198) 55.5 ± 1.6 33.7
Rigid (31 / 93) 55.5 ± 1.9 35.4
Non-rigid Solid (7 / 21) 68.1 ± 1.2 44.5
Fluid Dynamics (15 / 45) 57.6 ± 2.7 29.7
Optics (8 / 24) 41.0 ± 1.2 23.3
Thermodynamics (3 / 9) 67.3 ± 3.4 43.0
Magnetism (2 / 6) 37.0 ± 4.3 29.3
Rigid를 뺀 나머지 (35 / 105) 55.5 32.3

± 는 run 간 표준편차(ddof=1). Predict2.5는 1 run이라 SD가 없다. 장면 단위 bootstrap 95% CI는 scores/summary_mean_std.csv.

관찰 (해석은 §5 주의 참고)

  • 두 모델 모두 Optics가 가장 낮다(Nano 41.0, Predict2.5 23.3). 특히 Spatiotemporal IoU와 MSE가 낮다.
  • Rigid가 점수의 원천이라는 근거는 약하다. Nano는 Rigid(55.5) = Rigid 제외 나머지(55.5). Predict2.5는 Rigid(35.4)가 나머지(32.3)보다 약간 높지만, 장면 bootstrap CI가 크게 겹친다.
  • Fluid는 두 모델에서 다르게 나온다. Nano는 Rigid 수준(57.6)이지만 Predict2.5는 Rigid보다 낮다(29.7). Predict2.5는 Fluid의 Spatial·Weighted spatial IoU가 특히 낮다(29.6 / 17.2). 지표 4개별 표: figures/four_metrics_table_nano.png, figures/four_metrics_table_p25.png.

재현 확인: Nano 4 runs의 공식 점수가 공개 수치(55.51 ± 1.57)와 run별 최대 0.012점 차이로 일치한다(54.79 / 53.66 / 56.49 / 57.11).


2. 실험 설정 (비교용)

2.1 공통: 벤치마크와 채점

항목 값
벤치마크 Physics-IQ Verified (arXiv 2606.18943), GT = Anates-Labs-Research/Physics-IQ-Verified @ b65f567
트랙 V2V만
샘플 66 장면 × 3 시점(left / center / right), take-1 = 198개
채점기 google-deepmind/physics-IQ-benchmark @ 9e1f1e5 (physiq/run_physics_iq.py), 코드 수정 없음
점수 final_score_view = 샘플마다 지표 4개(Spatial IoU, Spatiotemporal IoU, Weighted spatial IoU, MSE)를 각각 [0,1]로 자른 뒤 평균, 그 198개의 평균
샘플별 점수 공식 IQTable.compute_scores_scenario_by_view() 출력을 그대로 사용(재구현 없음). 198개 평균 = 공식 final_score_view 일치를 assert로 검증
범주 공식 descriptions/best_practice/descriptions_base.csv의 category 열. 장면 수 38 / 15 / 8 / 3 / 2
Rigid Solid Mechanics 중 31개 장면(93 샘플). 팀의 rigid-93 목록(boxworld-core-20261101/physics_iq/frontend_v2/piq_v2b)과 동일. 잠정: 동료 목록과 최종 대조 필요
불확실성 run 간 SD + 장면 단위 bootstrap(B=5000, 같은 장면의 3 시점을 함께 재표집)
채점 FPS 각 모델이 생성한 FPS 그대로(Nano 24, Predict2.5 16). Verified GT에 24·16 FPS 테스트 영상과 실제 마스크가 들어 있어 보간 없음
채점 환경 Python 3.14.10, uv sync(pyproject), OpenCV는 같은 버전의 headless 빌드(서버에 libGL 없음), static ffprobe. env/scorer_pip_freeze.txt

2.2 Cosmos3 Nano (공개 영상, 재채점만)

항목 값
영상 출처 deepkyu/cosmos3-physicsiq-verified-v2v-prompt-upsampling @ 9e62b70, 폴더 cosmos3-nano-v2v/run_0..3
제출 카드 기재 설정 cosmos3_8b(카드의 parameters 표기 16B), 1280×720, 24 fps, 189 frames 생성 후 앞 69 frames 잘라 5 s(120 frames), UniPC 50 steps, guidance 3.0, shift 4.75, seeds 0–3, BoN 없음
프롬프트 GPT-6-Astra 프롬프트 업샘플링(카드의 descriptions.csv). 업샘플링 없는 기본 Nano와 다름

2.3 Cosmos-Predict2.5-2B (직접 생성)

팀 gen_cosmos의 vanilla baseline 설정을 그대로 썼다(code/gen_cosmos/, 원본 대비 변경은 경로 7줄뿐: CHANGES_vs_team_gen_cosmos.diff).

항목 값
체크포인트 nvidia/Cosmos-Predict2.5-2B, revision diffusers/base/post-trained (snapshot 0d37c74), LoRA 없음
텍스트 인코더 / VAE nvidia/Cosmos-Reason1-7B @ 3210bec (임베딩 미리 계산) / Wan2.1 VAE
런타임 diffusers 0.39.0, torch 2.5.1+cu121, transformers 4.57.6 (env/generation_pip_freeze.txt)
해상도 / 길이 832×480, 93 frames @ 16 fps
조건 conditioning 영상(30 fps)의 마지막 13 frames를 16 fps로 재샘플(switch frame에서 끝남) → 4 latent frames
출력 조건 13 frames를 뺀 80 frames = 5 s
샘플러 UniPC(공식 scheduler config), 35 steps, CFG 7, 기본 negative prompt
프롬프트 Physics-IQ Verified base 템플릿(descriptions_base.csv의 description, 업샘플링 없음)
seed / runs seed 0, 1 run (BoN 없음)
Guardrail off (가중치 접근 불가. 출력을 바꾸지 않고 차단만 함. 팀 baseline은 on)
생성 시간 영상당 약 10.4분, 최대 VRAM 8.6 GB (RTX 3090)

2.4 두 모델 사이에 다른 점 (해석 주의)

프롬프트(Astra 업샘플링 vs base), FPS(24 vs 16), 해상도(720p vs 480p), run 수(4 vs 1), 모델 크기가 모두 다르다. 두 모델의 절대 점수 차이를 모델 성능 차이로 읽으면 안 되고, 범주 간 상대 패턴을 보는 용도다.


3. 재현 방법

# 0) 평가 코드와 GT
git clone https://github.com/google-deepmind/physics-IQ-benchmark && cd physics-IQ-benchmark && git checkout 9e1f1e5
uv sync
hf download Anates-Labs-Research/Physics-IQ-Verified --repo-type dataset --exclude 'full-videos/take-2/*' \
  --local-dir <BASE>/physics-IQ-benchmark-verified          # 접근 동의 필요

# 1) Cosmos3 Nano: 공개 영상 받아 run마다 채점 (code/pipelines/nano_score.sh)
hf download deepkyu/cosmos3-physicsiq-verified-v2v-prompt-upsampling --repo-type dataset \
  --revision 9e62b70f490817a8132f331cf6de027c023f3cae --include 'cosmos3-nano-v2v/*' --local-dir nano
python physiq/run_physics_iq.py --input_folders nano_run_0 --output_folder out/nano_run_0 \
  --descriptions_file nano/cosmos3-nano-v2v/descriptions.csv --benchmark_base_folder <BASE> --n_process 16
#   (입력 폴더 이름은 run마다 달라야 한다: 채점기가 생성 영상 마스크를 폴더 이름으로 저장하고 파일 이름도 바꾼다)

# 2) Cosmos-Predict2.5-2B: 생성 -> 채점 (code/pipelines/p25_pipeline.sh)
#    프롬프트 임베딩 (Reason1-7B, GPU 1장)
python code/gen_cosmos/scripts/encode_prompts_baseline.py all
#    GPU마다 worker 1개 (잠금 파일 기반 작업 큐, 재실행하면 이어서 진행)
CUDA_VISIBLE_DEVICES=$g python code/gen_cosmos/scripts/run_baseline.py --subset all --modes v2v --guardrail off --out <OUT>
#    (기본값: --height 480 --width 832 --steps 35 --guidance 7.0 --seed 0 --v2v-cond-px 13)
python physiq/run_physics_iq.py --input_folders <OUT>/v2v --output_folder out/p25_run_0 \
  --descriptions_file descriptions/best_practice/descriptions_base.csv --benchmark_base_folder <BASE> --n_process 16

# 3) 범주별 집계와 그림
python code/analysis/aggregate.py --results <run CSVs: nano_run_k.csv, p25_run_0.csv> --models nano p25 --out scores
python code/analysis/metric_table.py scores/summary_mean_std.csv p25 four_metrics_table_p25.png
python code/analysis/metric_panels.py scores/per_sample_scores.csv metrics_by_category_panels.png

환경 메모(이 서버 한정): driver 525에서 torch cu121이 CUDA error 804를 내면 525용 libcuda.so만 담은 폴더를 LD_LIBRARY_PATH 앞에 둔다. 대량 I/O는 NFS가 아니라 로컬 디스크에서 했다(NFS에서 채점기 영상 읽기가 타임아웃났음).


4. 파일 구성

경로 내용
cosmos-predict2.5-2b/videos/*.mp4 Predict2.5-2B 생성 영상 198개(생성 구간 5 s만, 16 fps, 832×480). 파일 이름 = 공식 generated_video_name
cosmos-predict2.5-2b/logs/ 생성 설정(config.json)과 영상별 로그(시간, VRAM, GPU)
scores/metadata_198.csv 샘플 ID ↔ 장면, 시점, 범주, rigid 여부
scores/per_sample_scores.csv 모델 × run × 198 샘플의 Verified 점수와 지표 4개
scores/category_scores.csv, scores/summary_mean_std.csv run별·범주별 점수, run 간 평균 ± SD, bootstrap CI
scores/official/ 공식 채점기 원본 출력(run별 CSV, *_metrics.json)
scores/nano_original_physics_iq/ (참고) Nano를 원본 Physics-IQ 방식(원본 GT + 원본 공식)으로 채점한 결과: 58.84 ± 1.94
figures/ 표·그래프 PNG
code/ 생성(gen_cosmos), 채점·집계(pipelines, analysis) 코드와 입력 프롬프트 목록
env/ 패키지 버전, 모델·데이터 revision

포함하지 않은 것: Physics-IQ GT 영상과 마스크(접근 동의 필요 데이터), Cosmos3 Nano 영상(원본: deepkyu/cosmos3-physicsiq-verified-v2v-prompt-upsampling @ 9e62b70).


5. 해석 시 주의

  • Physics-IQ 점수는 "물리가 맞았는지"가 아니라 "움직임이 GT와 같은 곳·같은 때에 일어났는지"(motion mask 기반)를 잰다. 범주 간 차이에는 현상 난이도와 움직임의 크기·속도가 섞여 있다.
  • 범주 이름과 실제 운동이 꼭 일치하지 않는다. 예: Optics의 mirror-ball-fall(Nano 96.7)은 사실상 공 낙하다. 범주 안 편차가 범주 간 차이보다 크다.
  • Thermodynamics(3 장면), Magnetism(2 장면)은 비교하기에 너무 작다. Optics(8 장면)도 CI가 넓다.
  • 같은 Nano라도 출처마다 점수가 다르다: Verified + Astra 55.5(이 저장소), README Verified 리더보드 기본 Nano 43.0, Cosmos 3 보고서 Table 13(원본 채점) 50.2. 인용할 때는 채점 방식과 프롬프트 설정을 함께 적는다.
Downloads last month
78