Hi Topaz Labs,
This is a short technical report:
Optimal SD > 4K workflow using Proteus Standard + Second Enhancement
After extensive testing on version 1.4.0 of Topaz Video AI, I would like to share several findings regarding image quality, model behavior, and performance when processing 576i SD sources into 4K using the new Second Enhancement feature.
These observations were made on an MSI Titan 18HX equipped with an Intel Ultra 9 285HX and an NVIDIA RTX 50‑series GPU.
Key discovery:
- Second Enhancement produces superior quality with only one generation;
- Enabling Show Second Enhancement Feature in the Preferences unlocks a two‑pass enhancement pipeline within a single export, avoiding an intermediate encode.
This provides two major benefits:
- No generational loss (only one AV1 encode at the end);
- Higher quality, especially for SD interlaced sources.
This feature is extremely powerful and should be highlighted more prominently for advanced users.
Optimal two‑pass workflow for SD 576i > 4K
Pass 1:
- Output Resolution: 3840×2160 (final resolution, not intermediate);
- Pixel type: Square pixel, Force anamorphic contents to be square pixel;
- Crop: Letterbox/Pillarbox (removes overscan that causes hallucinations and slows inference);
- Proteus Interlaced (cleaning + anti‑banding + 1080p intermediate)
- AI Model: Proteus (Interlaced variant);
- Add Noise: 3, Noise=3 removes banding and MPEG artifacts without leaving visible noise after Pass 2;
- Recover Detail: 0, Prevents halos, ringing, and compression defects from being amplified;
- Grain: Off;
- Focus Fix: Off;
- Telecine: Off;
- Field Order: Auto‑Detect;
- Advanced Tuning: Automatic.
Why this works:
Pass 1 performs a clean, stable deinterlace, removes banding via controlled dithering, and prepares a clean 1080p intermediate for the second pass.
Pass 2:
- Second Enhancement: Enabled;
- Proteus Progressive (detail reconstruction + final 4K upscale);
- Intermediate: Auto (Topaz splits the upscale 50%/50%);
- AI Model: Proteus;
- Model Variant: Proteus Standard, More detailed and sharper than Proteus Natural for this type of content;
- Recover Detail: 20, Ideal balance: sharp 4K reconstruction without halos;
- Add Noise: 0;
- Advanced Tuning: Automatic.
Why this works:
Pass 2 removes the noise added in Pass 1, reconstructs fine details, and completes the upscale to 4K with a natural, clean look.
Performance results on MSI Titan 18HX AI A2XWJG
With this two‑pass workflow:
- Speed: 7.9–8.0 fps (after CPU stabilization);
- CPU usage: 52–60%;
- CPU temperature: ~87°C
- GPU usage: 39–43%
- GPU temperature: ~65°C
System remains fully responsive (video editing, photo editing, browsing, encoding possible simultaneously).
This demonstrates that the Second Enhancement pipeline is fully stable and efficient on high‑end hardware.
Why this workflow outperforms single‑pass or manual approaches
- Noise=3 in Pass 1 eliminates banding before detail reconstruction;
- Recover Detail=20 in Pass 2 avoids amplifying SD compression artifacts;
- Auto intermediate scaling ensures balanced enhancement;
- Automatic tuning now avoids over‑sharpening (fixed since 1.3.x);
- Only one final encode → no generational loss;
- Proteus Interlaced + Proteus Progressive is currently the highest‑quality combination for SD interlaced sources;
- On a 77" Samsung QD‑OLED, the improvement is clearly visible: clean gradients, stable textures, no halos, no ringing, and excellent 4K detail.
Suggestions for future versions
- Make Second Enhancement more visible for advanced users;
- Document the existence of the Interlaced variant of Proteus;
- Provide guidance on optimal Noise/Recover settings for SD sources
Consider factory presets for multi‑pass workflows (SD > HD > 4K)
Conclusion
This two‑pass Proteus workflow produces the best SD > 4K quality I have achieved so far, with excellent performance and stability.
The Second Enhancement feature is a major step forward and deserves more visibility.
Thank you for your continued work, this version is a significant improvement.
Best regard, Vincent.
