Through the analysis of various AI models, Nyx consistently stood out as the model with the least distortion and the best overall results. However, relying on a single model rarely delivered completely satisfying results. Combining multiple models to achieve optimal outcomes was time-consuming and cumbersome.
Model Observations
Iris Model performs excellently but is still not perfect. Some results appear slightly unnatural, possibly due to a lack of contextual understanding of human facial features.
Rhea Model works well for medium-quality videos but appears overly artificial upon zooming in, likely due to excessive corrections. XL models also exhibit similar issues, and for now, they seem more suitable for high-quality videos until v2 is released.
Frame Interpolation with Apollo, Chronos, Aion still has noticeable stuttering issues even after extensive adjustment of settings. When the original video has slight stutters, interpolation often does not result in significant improvement.
Estimate Feature Expansion
Expanding the Estimate feature to allow the AI to automatically determine the best model combination based on user settings, such as resolution (FHD or 4K), or the maximum or user-specified file size, would be incredibly useful. For example:
For low-quality videos: Use Theia/Gaia for detail recovery → Nyx for upscaling → Iris for facial refinement.
For high-quality videos: Use Nyx/Iris for refinement → Proteus/Rhea/Artemis for final enhancement.
This idea has already been shared with the Topaz team, and I’d love to hear their thoughts on it.
Subject-Only Upscaling
A feature that allows the AI to track and upscale only a specific subject while removing the surrounding background would be incredibly helpful. I’m emphasizing this for the second time, but I strongly hope it will be seriously considered.
