arXiv 2026

SNF-Bench:
Separating Static Drift from Natural Flow in Long-Horizon Fixed-Camera Video Generation

Matiur Rahman Minar1, Seunghun Oh2, Ganghyeon Jeong2, Unsang Park1,2
1Department of Computer Science and Engineering, Sogang University 2Department of Artificial Intelligence, Sogang University
Paper Soon Code BibTeX
Long-horizon fixed-camera nature video examples across sea, snow, river, and storm scenes
SNF-Bench evaluates long-horizon fixed-camera video generation by separating what should stay fixed from what should keep moving naturally.

TL;DR

Long fixed-camera videos fail in two different ways: backgrounds slowly drift, and natural motion fades, jitters, or leaks into static regions. SNF-Bench partitions each scene into static support and natural flow, then reports static fidelity, flow persistence with absolute magnitude, and drift leakage separately.

S Static support Background regions that should preserve identity and layout.
N Natural flow Water, rain, snow, fire, clouds, foliage, and other persistent motion.
F Frame horizon Long rollouts where errors compound over time.
3 Metric families Static fidelity, flow persistence, and leakage diagnostics.

Benchmark Focus

Fixed-camera generation is a special stress test for video models. The camera cannot hide geometry drift with viewpoint motion, while natural elements must remain lively over long durations. A single aggregate quality score can blur these failure modes together, making models look good even when the background slides or the flow collapses.

SNF-Bench makes those failures explicit by evaluating static and dynamic scene regions with separate diagnostics. The goal is not just to ask whether a generated video looks plausible, but to ask where the plausibility comes from and where it breaks.

1. Partition

Separate each scene into static support and natural-flow regions so that the benchmark can score the right behavior in the right place.

2. Track

Measure how the generated sequence evolves over time, emphasizing long-horizon accumulation rather than only early-frame quality.

3. Disentangle

Report static drift, flow persistence, flow magnitude, and leakage as distinct signals instead of folding them into one opaque score.

4. Compare

Use the decomposed scores to compare long-video generators under the same fixed-camera assumptions.

Metrics

SNF-Bench is organized around the specific failure modes that appear in long fixed-camera generation.

Static fidelity
Measures whether static support preserves the original scene identity, structure, and layout across the rollout.
Flow persistence
Measures whether dynamic regions continue to move naturally instead of freezing, pulsing, or collapsing over time.
Flow magnitude
Accounts for absolute motion strength so a model is not rewarded for trivial low-motion outputs.
Drift leakage
Detects motion spilling into regions that should remain fixed, a common long-horizon artifact in static-camera videos.

Code And Data

The benchmark implementation is available at github.com/minar09/snf-bench. Paper and dataset links will be added here when public release artifacts are ready.

BibTeX

@article{minar2026snfbench,
  title   = {SNF-Bench: Separating Static Drift from Natural Flow in Long-Horizon Fixed-Camera Video Generation},
  author  = {Minar, Matiur Rahman and Oh, Seunghun and Jeong, Ganghyeon and Park, Unsang},
  journal = {arXiv preprint},
  year    = {2026}
}