Evidence & limits
Images in STEADY
- Sample portrait
- A fictional, AI-generated adult.
- Previews
- YouCam renders; the sample reuses recorded ones.
- Sample checkpoints
- Recorded YouCam images represent simulated makeup passes.
- Your photos
- Private; live video stays in your browser.
What a checkpoint measures
STEADY compares regional image color against your baseline and chosen target. Untouched skin patches correct global exposure and color changes. A separate patch checks for inconsistent lighting. Progress measures image color, never product quantity, beauty, or completed application.
Unreliable tracking, pose, visibility, or reference lighting can prevent measurement. Region-only lighting changes can still resemble pigment. Camera processing, skin appearance, texture, and products also affect readings. These engineering thresholds aren’t calibrated cosmetic assessments. Reloading clears local frames; STEADY asks before capturing another baseline.
What the sample demonstrates
Recorded images use the same pixel calculation as local checkpoints. The guide comes from the actual local MediaPipe model. This demonstrates target choice, measurement, and method adaptation. It does not demonstrate a person applying makeup. Historical sessions retain their original targets and measurement versions.
Methods and decisions
The planner checks confirmed movements, products, and tools against finite methods. Adaptation preserves confirmed steps and prefers the existing product family. Instructions are AI-assisted drafts, without makeup-artist or participant validation. Follow product directions; skip or stop movements you don’t want. Light, Build, and Full are visual targets, not calibrated doses. Your chosen photo and assessment describe your physical result.
What leaves your device
Live frames, baselines, checkpoint pixels, and face coordinates stay locally. Session history saves measurements, target identity, and your choices. Chosen starting photos and requested previews use private server storage. Consented second-opinion stills reach Gemini or OpenAI with measured context. Their suggestions cannot advance a step. Enabled voice controls send short recordings to Deepgram. Deletion queues private photos for removal; provider retention remains separate.
How we evaluate
Software tests, synthetic experiments, provider renders, and browser checks stay separate. No physical-application outcome, participant benefit, or time saving is established. Reproducible methods and actual denominators are published below.