FARTBench: Flatulatory Audiovisual Realism & Temporal Benchmark
A Standardized Framework for Evaluating Acoustic Fidelity, Temporal Localization, Anatomical Causality, Fabric Response, and Social Coherence in Generative Video
A deliberately over-engineered investigation into an audiovisual problem nobody asked to have benchmarked.
A note on seriousness
FARTBench uses real benchmark methodology to investigate an intentionally ridiculous test case. The measurements, blind evaluation, provenance, and reproducibility work are meant seriously. The subject matter is not.
Abstract
Generative video systems have advanced rapidly in instruction following, photorealism, motion synthesis, native audio generation, and long-horizon temporal coherence. Yet contemporary evaluation suites systematically under-characterize a deceptively difficult class of multimodal events: brief, body-localized, acoustically structured, socially consequential flatulatory emissions. Such events jointly stress temporal synchronization, source attribution, fine-grained body mechanics, deformable-fabric response, acoustic synthesis, contextual restraint, and causal reaction modeling. We introduce FARTBench, the Flatulatory Audiovisual Realism & Temporal Benchmark, a reproducible evaluation framework spanning six benchmark families, 1,000 canonical prompts, 200 adversarial scenarios, seven primary scoring dimensions, ten standardized failure codes, blinded human preference evaluation, and cost-normalized efficiency metrics. We define FARTScore, a weighted composite intended to reward perceptually convincing emissions while heavily penalizing source mislocalization, phantom audio, temporal inversion, anatomical catastrophe, social nonresponse, and unrequested defecation. Synthetic baseline experiments illustrate expected failure surfaces and demonstrate why seemingly trivial prompts can expose multimodal weaknesses obscured by conventional video benchmarks. We further introduce the Perceived Wetness Index (PWI), an intentionally low-weight exploratory metric whose methodological status remains controversial. FARTBench is released conceptually as a model-agnostic, vendor-neutral, reproducible benchmark for researchers unwilling to accept that this problem should remain unmeasured.
Executive finding
The central empirical hypothesis of FARTBench is simple: if a generative system cannot reliably determine who farted, when they farted, what it sounded like, how their clothing moved, and whether the surrounding humans noticed, then claims of robust audiovisual world modeling remain incomplete.
Keywords
- generative video
- audiovisual synchronization
- causal attribution
- embodied simulation
- acoustic realism
- deformable fabric
- social reasoning
- flatulence
- multimodal benchmarking
- human evaluation
Record
- Report number
- FB-TR-2026-001
- Version
- 1.0
- Date
- September 2026
- Author
- Cale Frombach
- Affiliation
- Independent researcher
- Length
- 35 pages
- Format
- SHA-256
- a6863fd94523d481d51f91c38e6cd78b2e1b8a71e251c13edaa34ecd80f4b12c
Not peer reviewed. No DOI is assigned, no journal has accepted it, and no institutional affiliation is claimed. It is a self-published technical report and is described as one.
Report and protocol are versioned separately
The report describes a broader conceptual benchmark than the one currently in operation: it discusses six benchmark families, a thousand canonical prompts, and ten failure codes. The operational protocol governing this site is narrower and newer.
| Aspect | Technical report v1.0 | Operational protocol Core R01 |
|---|---|---|
| Scope | Six benchmark families, 1,000 canonical prompts, 200 adversarial scenarios | 4 scenes × 16 configurations: 63 clips in blind review (1 cell refused by moderation) |
| Failure codes | Ten standardized codes | None. Scene-specific factual questions record whether the event happened and whether its source or cause was right. |
| Human evaluation | Blinded human preference evaluation | 4 Official reviewers per clip: 252 blind evaluations, each locked. Community review is separate and never changes the Official results. |
| Reported results | Synthetic baselines, explicitly labelled in the report | Core R01 results (Results V2), revealed and frozen; Overall Ranking V1, frozen. |
| Scoring dimensions | Seven primary dimensions | Six questions per clip: three scene-specific factual questions, then three ratings. Overall Ranking V1 uses the factual answers only. |
Where the two disagree, the operational protocol governs this website. The report is the origin document, not the current specification.
Before Core R01, the pilot-era protocol Core v0.1.0 had 8 core tests, 15 failure codes (F01–F15) and seven weighted scoring dimensions (A T L P I S C). It remains documented, labelled legacy, on the original methodology page in the Archive.
Citation
Plain text
Frombach, C. (2026). FARTBench: Flatulatory Audiovisual Realism & Temporal Benchmark. FARTBench Technical Report FB-TR-2026-001, Version 1.0.
BibTeX
@misc{frombach2026fartbench,
title = {FARTBench: Flatulatory Audiovisual Realism & Temporal Benchmark},
author = {Frombach, Cale},
year = {2026},
number = {FB-TR-2026-001},
note = {FARTBench Technical Report, Version 1.0}
} Cite the report for the benchmark's conception. Cite the protocol version and prompt-pack version for anything measured — a score is meaningless without the protocol that produced it.