Bob and Luna put avatar and lip-sync models to a five-second test. Watch the results, compare the numbers, and check each model's inputs.
SUBJECT 01 Bob
SUBJECT 02 Luna01 / THE COMPARISON
Each result uses Bob or Luna's portrait or source video. Audio-driven models use their recordings; text-driven models use synthesized speech and are labeled. Open a result for the exact call and measurements.
Loading benchmark results…
02 / EXPERIMENTS & TRIAGE
Agent experiments, retries, and investigations stay separate from the benchmark pairs. Each report records the model, version when known, dates, method, evidence, and outcome.
Loading experiment reports…
03 / THE LANDSCAPE
We reviewed avatar and lip-sync endpoints from Fal, Atmee, and Reactor. Selected dollar rates are below 10¢ per second; Atmee's account rate is shown in credits.
04 / THE METHOD
Audio-driven models use Bob and Luna's five-second recordings. Reactor's text-driven models synthesize a new voice; their results are labeled and are not an exact audio comparison. Portrait models receive each character's production portrait.
Elapsed time includes setup or queue time. Reactor clips are captured from live streams, so their preparation and capture differ from queued file generation. Resolution and duration come from the archived output when available.
Fal estimates use the published rate and output duration. Reactor charges for reserved session time, including setup and idle time, so the final cost exceeds the rate times five seconds when the session takes longer. Actual charges can differ. Quality ratings stay in this browser.