Help center

Unified Reframing update

This Help page explains what Unified Reframing is, why it exists, and how it behaves end-to-end in ViralSlice. It covers the technical flow, layout logic, reliability guardrails, and where to find related operational details.

Instead of one fixed crop strategy, ViralSlice classifies clip context first and then applies the most suitable vertical layout behavior per content pattern.

How the pipeline works

Each generated clip goes through four reframing stages:

• Global scene analysis
• Timeline event detection
• Layout plan generation
• Renderer application with fallback guards

Newer builds also run vision-confidence checks and post-render framing QA so poor framing can auto-rerender with safer crop behavior before export.

Layout modes by content category

• Talking head: full-crop 9:16
• Podcast/interview: split + optional solo-focus windows
• Vlog/landscape: canvas layout with optional focus windows
• Tutorial/demo: top-align 9:16 layout

Advanced framing safeguards (new)

• Vision confidence pass: estimates subject center-bias and side-bar risk before final mode application.
• Talking-head override: when visual confidence strongly indicates off-center speaker framing, the planner can bias back to full-crop talking-head mode.
• Post-render visual QA: detects poor framing outcomes (for example severe side bars or edge-biased subject) and can auto-rerender with smart crop fallback.
• BlazeFace-first smart crop: we center the horizontal crop on detected faces across sampled frames when possible; letterbox clamp and motion/heuristic fallbacks apply when faces are not detected reliably (e.g. extreme angles or occlusion). Post-render QA may still rerun the crop when output framing looks unsafe.

Podcast dynamic switching (improved)

Podcast split/solo switching now uses stricter anti-jitter controls for more stable output:

• minimum sustained speaker-run duration before switching to solo
• minimum confidence score gates for solo/split transitions
• minimum window length to avoid tiny flicker segments
• switch cooldown to prevent rapid flip-flop behavior

Safety and reliability model

Dynamic behavior is guarded by multiple checks. If a mode-specific render fails validation, the pipeline falls back to smart reframing so clips still export reliably.

• Window normalization and overlap trimming
• Max window caps to avoid jittery switching
• Output sanity checks (duration drift and audio parity)
• Post-render visual QA checks before final export
• Fallback rendering when quality/safety checks fail

Rollout and tuning

Unified reframing is controlled by feature flags and tuned progressively in production. Recommended order is baseline first, then dynamic podcast, then dynamic vlog after stable telemetry.

Admin telemetry tracks warn rate, fallback rate, dynamic usage, and mode / renderer distribution for each window.

Related docs and support

For related product behavior and policy details: