I thought I would get better results from VG16, but it feels like each got its pros & cons
With VG16 the moire pattern is a lot less noticeable when training at high values (like 50%) or sometimes even gets completely trained out with lower values (even when zooming in to 200%)
With Alex I get a lot more sharpness, but the pattern stays visible when zooming in even when training at very low values (also much faster training of course)
I tried squeeze too, but didn't let it run for very long because it just seemed way worse than both alex & vg16 in the early stages
This is with SSIM as the main loss function btw and no other extra loss functions
What are other people's experiences with these loss functions?
What I've started doing on my latest model is using a combination of alex & vg16 and I'm really liking the results so far (100 SSIM / 5 lpips_alex / 30 lpips_vg16 / 90 FFL)
--> it feels like adding 5% alex rly improves the sharpness compared to just using vg16 - not sure if it's worth the heavier GPU load though, might be better to tweak my phaze-a settings a bit to bring out more detail instead of combining 2 lpips loss functions
Last edited by d3x on Fri Apr 05, 2024 7:22 pm, edited 8 times in total.
You're seeing VGG16 provide smoother results with fewer moiré patterns, but less sharpness compared to AlexNet, which gives sharper images but retains more artifacts. SqueezeNet is lightweight but less effective early on. By combining AlexNet and VGG16 with SSIM, LPIPS, and FFL, you're balancing sharpness and artifact reduction, though it adds GPU load. Tweaking phaze-a settingscould enhance detail without needing both LPIPS models. Your current combination seems effective, but fine-tuning the weights or settings may optimize performance further.
I thought I would get better results from VG16, but it feels like each got its pros & cons
With VG16 the moire pattern is a lot less noticeable when training at high values (like 50%) or sometimes even gets completely trained out with lower values (even when zooming in to 200%)
With Alex I get a lot more sharpness, but the pattern stays visible when zooming in even when training at very low values (also much faster training of course)
I tried squeeze too, but didn't let it run for very long because it just seemed way worse than both alex & vg16 in the early stages
This is with SSIM as the main loss function btw and no other extra loss functions
What are other people's experiences with these loss functions? scratch geometry dash
What I've started doing on my latest model is using a combination of alex & vg16 and I'm really liking the results so far (100 SSIM / 5 lpips_alex / 30 lpips_vg16 / 90 FFL)
--> it feels like adding 5% alex rly improves the sharpness compared to just using vg16 - not sure if it's worth the heavier GPU load though, might be better to tweak my phaze-a settings a bit to bring out more detail instead of combining 2 lpips loss functions
VG16 vs. AlexNet
VG16 (VGG16): This is a deeper network compared to AlexNet, and it's known for capturing more nuanced features, which might be why you see a reduction in moiré patterns. However, it can be less sharp in the output because it might prioritize overall texture and pattern consistency over edge sharpness.
AlexNet: Being a shallower network, AlexNet tends to emphasize sharper edges and clearer features. This could explain why your results appear sharper but also why the moiré patterns are more persistent—AlexNet might not be as effective in learning or smoothing out those fine patterns.
Combining AlexNet (for sharpness) and VG16 (for reducing moiré patterns) is a good approach, balancing detail and artifact reduction. Fine-tuning the mix or trying different loss functions can help improve results while managing GPU load. drift hunters
