The replacement scene looked coherent, but the subject kept flat neutral source lighting against the moody studio backdrop, with no grounding shadow and a monitor element colliding with the hair silhouette.
What was measured
Lighting adaptation
Does the replaced background look naturally lit relative to the subject, or obviously composited?
decisive for this rankingtransformation
If the replaced background does not match the subject’s lighting, the composite looks fake, so this is central to replacement quality. (3 of 3 judges)
What was given, what came back
Test input: Indoor Talking Head · video · group: remove-or-replace-video-backgrounds-using-ai
Input — what we sent
Indoor Talking Head
An indoor talking-head video of a person speaking directly to the camera with a static background, natural hand gestures, and facial expressions. It was used to test how well tools preserve the subject while replacing an indoor background.
Why this input is hard
- · Static camera performance
- · Face and body segmentation
- · Facial expression preservation
- · Background replacement accuracy in an indoor environment
Output — unretouched



Also checked on this input — same tool, 1 other criterion
Provenance
- Observation
- 2176350b-7bf3-4457-a943-e84786bd3cc9
- Evidence run
- 10155228-7ce8-438a-a210-547331ef080b
- Study
- Remove or Replace Video Backgrounds Using AI
- Research task
- 86ba42c2d
- Tested at
- not recorded
- Source
- first-party
- Evidence state
- verified
- Proof shown
- input + output shown
- Cost / latency
- not captured
- Repeat run
- not captured
- Tester
- not captured
The last three rows are honest blanks, not placeholders — our capture has no field for them yet.
Query this
get_evidence({
tool: "media-io",
scenario: "remove-or-replace-video-backgrounds-using-ai"
})MCP · mcp.aidemos.com/api/mcp
Free with attribution.
Same input, same check — 1 other tool
measured on Lighting adaptation
This evidence is published in
Real inputs and real outputs, no retouching · every cell queryable via API & MCP · aidemos.com