SPLICE

Cross-shot visual consistency detection for narrative film and AI-generated video. CS231n Spring 2026 final project.

Team: Devon Smith, Lily Bailey, Xander Hnasko

What we're building

A model that takes the pair of frames at a shot-boundary cut — the last keyframe of one shot and the first keyframe of the next — and outputs a single continuity score: how consistent the visual change across the cut is with variation that plausibly occurs within one continuous scene.

Motivating use case: AI video tools (Runway, Sora, Luma, Pika) generate clips independently with no cross-shot continuity, so assembled multi-shot sequences show visible jumps at cuts (lighting, colour, background). SPLICE flags those for human review. It is a cinematographic-similarity scorer trained on real film — not an AI-gen detector; whether a clip is AI-generated is deliberately outside what the model sees.

See reports/project_spec.md for the architecture and reports/related_work.md for the literature review.

Versions

• v0 (current milestone) — frozen DINOv2 ViT-B/14 + logistic regression on a 2305-d pair feature, against three baselines (HSV chi-square, CLIP cosine, raw DINOv2 cosine), with an inference threshold calibrated from within-shot variation. Results: reports/v0_results.md.
• v1 — MLP head and ablations (boundary vs mean-pooled frames, negative mining).
• v2 — supervised contrastive loss, partial fine-tuning, AI-generated eval set.

Data

MovieNet, 318-movie scene-segmentation subset (BaSSL annotations: per-shot invideo_scene_id / boundary_label). Keyframes come from the HuggingFace mirror ZhengPeng7/MovieNet — the original BaSSL Aliyun OSS link is dead (HTTP 404). Large data lives under /mnt/disks/splice-data on the GCP VM, not in git.

Pipeline

micromamba activate consistency
python scripts/prep/build_cut_index.py        # labelled cut index (Parquet)
python scripts/prep/embed_keyframes.py        # cache DINOv2 embeddings (GPU)
python scripts/prep/build_pair_features.py --mode boundary   # 2305-d pair features
python scripts/train/v0_logistic.py           # train v0 + baselines
python scripts/eval/calibrate_threshold.py    # within-shot threshold calibration

scripts/inspect_movie.py --imdb_id <id> is a diagnostic for the annotation format.

Repo structure

• configs/ — YAML hyperparameter configs (v0_default.yaml)
• scripts/ — runnable entry points (prep/, train/, eval/)
• src/ — importable modules (data/, models/, eval/, losses/)
• tests/ — unit tests; run with pytest tests/
• reports/ — spec, related work, milestone results