Viralance: protocol for a future AI video comparison
Prepared 2026-09-18. This is a proposed method and blank worksheet, not new results.
Historical data: https://viralance.com/research/ai-video-model-benchmark-2026

QUESTIONS
1. Does a product label remain readable and correct?
2. Does a single subject preserve its shape during motion?
3. Does the delivered file match the requested format, duration and audio setting?

PRE-REGISTER BEFORE SPENDING
Record the maximum total budget, selected models/endpoints/versions, supported settings,
three source scenes (product label, object motion, landscape), three attempts per
model/scene, and an acceptance rubric. Use licensed or owned sources. Save source
hashes, exact prompts and requests; retain random seeds when supported. Do not use
private customer media. Report unsupported configurations separately, not as failures.

FAIR COMPARISON
Hold source, prompt and compatible settings constant. Randomize run order. Do not
silently substitute another model or discard failed runs. Publish all attempts and
state whether preparation cost, retries and refunds are included. Use the actual
pack rate paid, not a hypothetical cheapest rate. Stop at the declared budget.

MEASUREMENT
Probe every downloaded output with ffprobe and retain the JSON and file hash.
Example: ffprobe -v quiet -show_format -show_streams -of json output.mp4
Measure dimensions, duration and audio-track presence separately from visual quality.
Record submission/completion timestamps and failures. An audio stream is not proof
of meaningful sound. Preserve provider errors without credentials or user identifiers.

REVIEW RUBRIC
Review start, middle, end and the full clip. Record label spelling, shape consistency,
framing, motion and visible defects. Two reviewers should score anonymized outputs
independently where feasible. Record disagreements. No sales or ranking claim can be
inferred from this generation test. Report sample size and limitations alongside results.

BLANK WORKSHEET (one row per attempt)
run_id,test_date,model,endpoint_version,source_hash,prompt,settings,seed,attempt,
submitted_at,completed_at,status,error_category,credits_deducted,credits_refunded,
actual_pack_usd_per_credit,output_hash,width,height,seconds,audio_track_present,
label_correct,shape_consistent,framing_correct,motion_matches,accepted,reviewer,notes

PUBLICATION
Publish protocol, licensed inputs, prompts, request settings, output evidence and
complete results together. Separate historical observations from current quotes.
Record corrections and actual experiment dates. Aggregate costs and failures honestly.