Upload README.md

README.md

@@ -0,0 +1,134 @@
+# Matrix-Game 2.0 Self-Consistency Reproduction
+
+This repository reproduces the Matrix-Game 2.0 self-consistency (SC) row in
+*Reasoning-Structured Videos: A Stratified Diagnostic Suite for Compositional
+Consistency in World Models*.
+
+It evaluates the released full-rollout model videos directly. Ground-truth videos
+are not needed for SC.
+
+## What is reproduced
+
+The statistical unit is one relation graph:
+
+- **Inverse-SC:** generated first frame versus generated final frame.
+- **Loop-SC:** generated first frame versus generated final frame.
+- **Equivalence-SC:** generated final frame of branch A versus generated final
+  frame of branch B.
+
+The release contains 448 Inverse graphs, 445 Loop graphs, and 239 Equivalence A/B
+pairs. The script checks these counts and pairing integrity before evaluation.
+
+## Data
+
+Reproduction bundle (videos, evaluator, and this README):
+
+<https://huggingface.co/datasets/VideoWorldmodel/Evaluation>
+
+The generated videos were originally released at:
+
+<https://huggingface.co/datasets/WhynotGPT/Nips_WM_Eval_qzf>
+
+The benchmark GT dataset is available separately, but is not required by this SC
+reproduction:
+
+<https://huggingface.co/datasets/VideoWorldmodel/ReasoningStructureTestset>
+
+Download and extract the MP4-only reproduction archive:
+
+```bash
+python -m pip install -U huggingface_hub
+hf download VideoWorldmodel/Evaluation \
+  --repo-type dataset \
+  --include "MatrixGame2_SC_videos.zip" \
+  --local-dir .
+unzip MatrixGame2_SC_videos.zip -d data/Nips_WM_Eval_qzf
+```
+
+The expected layout is:
+
+```text
+data/Nips_WM_Eval_qzf/
+|-- inverse_easy/      # 246 MP4s
+|-- inverse_hard/      # 202 MP4s
+|-- loop_easy/         # 247 MP4s
+|-- loop_hard/         # 198 MP4s
+`-- equivalence/       # 478 MP4s = 239 A/B pairs
+```
+
+If a Git LFS clone was used, ensure the files are real video objects rather than
+small pointer files before evaluation.
+
+## Environment
+
+Python 3.10 or newer is recommended. For a standard CPU/default PyTorch install:
+
+```bash
+python -m venv .venv
+source .venv/bin/activate  # Windows: .venv\Scripts\activate
+python -m pip install --upgrade pip
+python -m pip install "numpy>=1.26,<3" "opencv-python-headless>=4.8,<5" \
+  "torch>=2.2,<3" "torchvision>=0.17,<1" "lpips==0.1.4"
+```
+
+For a specific CUDA build, install the matching `torch` and `torchvision` wheels
+from the official PyTorch instructions first, then install NumPy, OpenCV, and
+`lpips==0.1.4` from the command above without reinstalling PyTorch.
+
+The first LPIPS run downloads the official AlexNet weights.
+
+## Reproduce the paper row
+
+```bash
+python evaluate_matrix_game_sc.py \
+  --data data/Nips_WM_Eval_qzf \
+  --output results/matrix_game_sc \
+  --lpips \
+  --device auto \
+  --check-paper
+```
+
+Use `--device cuda`, `--device cuda:0`, or `--device cpu` to select a device
+explicitly. A fast PSNR-only audit can be run by omitting `--lpips` and
+`--check-paper`.
+
+Expected results:
+
+| Relation | Graph N | Recomputed LPIPS / PSNR | Paper LPIPS / PSNR |
+| --- | ---: | ---: | ---: |
+| Inverse-SC | 448 | 0.7061 / 10.4476 | 0.71 / 10.45 |
+| Loop-SC | 445 | 0.7173 / 10.6227 | 0.72 / 10.62 |
+| Equivalence-SC | 239 | 0.5918 / 12.5732 | 0.59 / 12.57 |
+
+`--check-paper` passes when all six values reproduce the precision reported in the
+paper. Two Equivalence pairs are exact pixel matches. Their PSNR is infinite, so
+they are reported in `psnr_exact_match_n` and excluded from the finite PSNR mean;
+this produces the reported Equivalence PSNR.
+
+## Outputs
+
+The output directory contains:
+
+```text
+per_graph.csv    # one row per Inverse/Loop graph or Equivalence A/B pair
+summary.csv      # means, standard deviations, and graph-bootstrap 95% CIs
+paper_check.csv  # computed values versus paper values at reported precision
+audit.json       # metric definitions, counts, versions, and bootstrap settings
+report.md        # compact Markdown result table
+```
+
+Confidence intervals use 10,000 graph-level bootstrap resamples with seed 2026.
+LPIPS uses `lpips==0.1.4`, AlexNet, and RGB inputs scaled to `[-1, 1]`. PSNR is
+computed on decoded RGB uint8 endpoints with `MAX=255`; no resizing is applied.
+
+## Tests
+
+```bash
+python -m unittest discover -s tests -v
+```
+
+## Scope
+
+This code reproduces the Matrix-Game 2.0 **SC endpoint metrics only**. It does not
+claim to reproduce the paper's GT-anchored metrics or FVD, which require additional
+generation-context and clip-sampling details.