Uniform prompt and reference input

A red ceramic coffee cup on a pale oak table beside a folded linen napkin. Keep the cup, logo, and handle shape stable while soft morning light crosses the table. Slow dolly-in, natural shallow depth of field.

Every candidate model receives this exact prompt and the same static product-photo reference file. The reference source URL and SHA-256 hash are recorded with the results.

Evaluation method

Run two unattended generations per model at 4 seconds and 720p, recording successes and failures.
Measure generation time from submit until the terminal event.
Record actual credits charged, not only advertised rates.
Assess product preservation, prompt adherence, motion stability, file integrity, and visible artifacts; do not invent numeric scores.
Observations are context-specific and are not universal rankings.

Result table

Run date	Model	Input	Output duration	Resolution	Generation time	Cost or credits	Observed strengths	Observed failures	Evaluator notes
No rows are published yet.

Candidate models and published rates

Agnes Video v2.0: 2 credits per second (published list rate).
Seedance 2.5: 480p 18 credits per second; 720p 39 credits per second (published list rates).
Kling v3 Omni: 39 credits per second without audio; 49 with audio (published list rates).
Veo 3.1: 4 seconds 80 credits (published list rate).

Published rates are not benchmark costs.