PathFinder-r/CortexVideo-results

CortexVideo Generated Video Results

CortexVideo: A Semantic-Spatial Dual-Anchor Framework for High-Fidelity fMRI-to-Video Reconstruction

Xiaoquan Shen, Jiaxuan Chen, Jiajun Li, and Gang Pan.

This release provides generated video results for three subjects from the CC2017 experiments. Training code, inference code, evaluation code, and model checkpoints are outside this initial results release.

Project repository on GitHub

The dataset viewer exposes the file manifest. Download the original GIFs using the commands below to inspect every animation frame.

Subject	Generation seed	Clips
1	111	1,200
2	0	1,200
3	1919	1,200

The dataset contains 3,600 original GIF files (1,581,226,028 bytes, approximately 1.58 GB). Clip IDs run from 0000 to 1199 within each subject.

File format

Each GIF displays ground truth on the left and reconstruction on the right, including display borders. The complete canvas is 518 × 260 pixels, with six frames and a total encoded playback duration of 720 ms. This GIF playback timing should not be interpreted as the original stimulus sampling rate. Files retain their original bytes; the release uses numeric filenames and records the original names in manifest.csv.

videos/
  subj_1/seed_111/0000.gif ... 1199.gif
  subj_2/seed_0/0000.gif ... 1199.gif
  subj_3/seed_1919/0000.gif ... 1199.gif
manifest.csv
SHA256SUMS
historical_metrics.json

Download and verify

python -m pip install huggingface_hub
hf download PathFinder-r/CortexVideo-results --repo-type dataset --local-dir downloads/CortexVideo-results
python downloads/CortexVideo-results/verify_download.py downloads/CortexVideo-results

To download only subject 2:

hf download PathFinder-r/CortexVideo-results --repo-type dataset --include 'videos/subj_2/**' --include manifest.csv --include SHA256SUMS --include historical_metrics.json --include README.md --include verify_download.py --local-dir downloads/CortexVideo-results
python downloads/CortexVideo-results/verify_download.py downloads/CortexVideo-results --subject 2

The verification script uses only the Python standard library and checks file sizes and SHA-256 hashes. SHA-256 establishes byte identity, not evaluation reproducibility.

Historical evaluation

Subject	2-way	50-way	CLIP-pcc
1	0.866533	0.261967	0.925596
2	0.876167	0.282808	0.936133
3	0.874483	0.280383	0.922854

These values were transcribed from the local evaluation logs associated with the selected generation seeds. They were not recomputed for this release, and do not exactly reproduce every manuscript table entry. Full precision and the logged uncertainty values are in historical_metrics.json.

The local classification evaluator uses MCG-NJU/videomae-base-finetuned-kinetics with six input frames and samples candidate classes with NumPy randomness over 100 trials per clip. Its evaluation random seed was not recorded; the generation seeds above do not fix this randomness. Re-evaluation may therefore yield different 2-way and 50-way scores. The logged classification ± values are averages of per-clip binomial standard errors, not standard deviations over repeated full evaluations.

For the historical input layout, split each decoded frame at column 259: the left half is ground truth and the right half is the prediction. This produces 260 × 259 panels and retains display borders. Cropping to 256 × 256, transcoding, changing frame selection, or changing model/preprocessing versions changes the evaluated input or protocol.

The current local CLIP-pcc implementation uses openai/clip-vit-large-patch14 to average cosine similarities over all distinct pairs of reconstructed frames, followed by averaging across clips. It does not compare reconstruction with ground truth in this metric. This differs from an adjacent-frame interpretation of temporal consistency. Its logged ± is the population standard deviation across clips. Exact model revisions and the complete historical evaluation environment were not captured in this results package.

Provenance and usage

These results come from CortexVideo experiments on CC2017. The ground-truth panels contain source stimulus content; they are not generated by CortexVideo. A related preprocessed dataset is available at gongzx/cc2017_dataset. This link provides context and does not establish a license for all source stimulus content. No blanket license for third-party stimulus content is granted by this repository.

When reporting evaluations, identify the dataset revision, subjects, generation seeds, file checksums, preprocessing, model revisions, and evaluation random seed. Cite the CortexVideo work by the title and authors above; a publication identifier will be added when available.