Why our quality tiers are speed tiers, not a quality ladder

Most enhancers imply a ladder: higher tier, better picture. On the AI Video Enhancer and quality-repair rails the English labels are Ultra Fast, Fast, General, High Quality, and Maximum — and we treat them as speed tiers. A slower pass spends more time. It does not promise a prettier frame on every clip. Download and compare 1s/3s/5s previews on your own damage before a long export.

• Damage is not one disease. Heavy compression and a clean-but-small source want different models
• A mid tier can win on blocky social re-encodes while a heavier pass barely moves a clean 1080p plate — and costs several times the minutes
• If we promised a monotonic quality ladder we would have to paper over the clips where it is false

Internal timing, as of July 2026, on an RTX 4090: six real degraded clips (480p–1080p mix), six frames per tier, mean of those frames, full desktop pipeline including inter-process overhead. Rounded, not a public benchmark you should treat as a contract:

• Ultra Fast: about 4 frames per second — a first look, or a weaker machine
• Fast: about 2.5–3 fps — the usual time/quality bargain; the Fast model is under 1 MB
• General: about 1 fps — the default for messy real-world encodes
• High Quality and Maximum: slower still. Use them for a final, then compare against General on the same 5s preview

Picture quality was not monotonic. Ultra Fast was close to heavier tiers on already-clean plates. General earned its keep on heavy compression and noise. That is why the promise we will keep is speed, not a beauty ranking. We still do not publish a VRAM floor (see download).

1. Start on General and preview 1s/3s/5s on the nastiest beat — guests can do this after download
2. If the source is already clean and only small, try Fast. You often save a lot of time with no visible loss
3. Weak GPU or a whole-timeline skim: Ultra Fast
4. Final delivery with time left: High Quality or Maximum, then A/B against General before you pay for export