Model Selection

Model Comparison

Decision Tree

1. Need text-guided object targeting? Use pro.
2. Portrait/face-cam content? Use human.
3. Need max throughput and simple scene? Use light.
4. Otherwise default to original.

A/B Testing Recommendation

For new domains, test at least 30 representative clips across two candidate models before rollout.

Quality Guardrails

• Prefer representative source clips when tuning model defaults.
• Keep input compression moderate to preserve edge details.
• Use explicit output format policy for transparent vs non-transparent workflows.
• Run A/B validation with the same clip set before changing defaults.

Implementation Checklist

1. Define payload schema validation in backend before forwarding requests.
2. Store model/output_format/background settings with each job record.
3. Add internal quality review for difficult scenes (hair, glass, motion blur).
4. Create runbook entries for model-specific failure cases.

Model Rollout Plan

1. Define scene buckets (portrait, complex object, simple bg).
2. Benchmark all 4 models on each bucket.
3. Set fallback order by scenario and latency SLO.
4. Deploy with model override switch in ops console.