Ordinary person
Body outline, fingers, occlusion and retained audio
Pending real source footage, processing and review.
AI Video Background Removal Tests
We are building a reproducible comparison of hair, fast motion, pet fur, low light and difficult backgrounds across Draft, Standard and Pro. The protocol below defines the evidence required; it is not a quality ranking.
Benchmark incomplete — verified results are pending. We will publish source clips, actual outputs and review notes before drawing comparisons. No processing-time averages, quality scores or winning model are claimed.
Which scenarios are being tested?
Fine hair
Flyaway detail and fringes over light and dark backgrounds
Pending real source footage, processing and review.
Fast motion
Motion blur, missing limbs and temporal flicker
Pending real source footage, processing and review.
Pet fur
Fine fur, paws and subject selection
Pending real source footage, processing and review.
Low light
Noise, dark edges and low subject/background contrast
Pending real source footage, processing and review.
Complex background
Similar colours, clutter and overlapping subjects
Pending real source footage, processing and review.
How will the comparison be reproduced?
- Use authorised footage and record its source, duration, frame rate, dimensions and SHA-256. Label camera footage and AI-generated material separately.
- Run the identical source through each quality tier with recorded model identifiers, settings, dates and output formats. Separate AI processing from preparation and export time; record the measurement boundary.
- Keep the original and actual output files. Inspect moving edges over light and dark backgrounds. Check audio sync and alpha in the recorded destination application/version.
- Record failures, missing details and limitations alongside successes. Repeated timings, hardware/network context and more clips are required before generalising a single result.
The three existing homepage demonstrations use AI-generated source clips and are illustrations, not this six-scenario, multi-tier benchmark. We have not repurposed them as measured scores.
What evidence will each result include?
Original and output downloads, model/version, quality, duration, dimensions, source and processed frame rate, elapsed time with its measurement method, credits, export format, audio/alpha checks, software version, reviewer, date, observations and limitations.