As an Amazon Associate we earn from qualifying purchases.
Anyone else running into inconsistent style applications when using StyleSynthesis in batch processing? I'm finding that even with identical input parameters and style references, the output styles can vary substantially from image to image. It's not a subtle difference either, sometimes one image will nail the style transfer perfectly, while the next will entirely miss the mark, almost as if it's ignoring certain style elements.
I've tried a few things like increasing the batch size (didn't help), normalizing the input images more aggressively (some improvement, but not a complete fix), and even splitting the batch into smaller chunks and processing them separately (still inconsistent). Wondering if it's a memory issue,or perhaps the stylesynthesis model is just inherently prone to this kind of variance.
Any tips, tricks, or solutions you've found to mitigate this issue would be greatly appreciated. Specifically, has anyone had success with pre-processing their style references in a certain way to make them more robust for batch applications, or tuning specific parameters in the StyleSynthesis algorithms (like content/style weights)? I'm using fairly standard convolutional neural network based StyleSynthesis, so experiences with that are particularly welcome.