Here’s a close-up to make it a little clearer.
Original image:
Processed with Recovery:
4 Likes
Thank you for added the corner to look out for - i almost missed it!
This definitely is something worth noting!
2 Likes
You do know that’s a 
check mechanism right?
Hi,
This is a known issue. We are working on fixing this.
2 Likes
Hmm… In images generated or upscaled by certain software, there’s often this really weird distortion that resembles the texture of “alligator skin”, particularly noticeable on fabric patterns. I could be mistaken, but I’m not sure whether the trained models have accurately learned the appearance of fabrics. For jeans, t-shirts, etc. - I sincerely hope that the team at Topaz is working on this issue.
This is a joke, right?
I only have this problem with Photo AI and Gigapixel AI. I don’t have it on Video AI.
Yes just joking. If it works with command line workaround then it must be a bug.
That’s great to hear. It’s pretty much the only issue Recovery has.
I agree, that sounds most probable and what I was suspecting. It’s good that you’re keeping an eye on this, and hopefully color prediction can be one additional metric to monitor during training. Might require a couple of good examples to be added to the (train/val/test) datasets though, or fiddling with weights when the loss determines which parameters should be prioritized for the (blunt instrument that is a) single score, as the only vehicle for steering the model towards learning what sort of reconstruction would be more faithful given the ground truth.
PS. and thanks for taking the time to respond, and with such detail Yuan!
1 Like
It looks like a tiling/stitching problem.
The rectangle I can discern is 448 pixels wide, which is a precise multiple of the Topaz model sizes (512, 192, 128 …). I can’t say exactly which models the app used for that image, but that the sharp color contrasts are clearly tiling edges is apparent, and that they are even multiples of the tiling models leads to that initial hypothesis. 448-192 = 256 = 2x128.
2 Likes
Thanks! I hope this is fixable?
If it is a tiling problem, then it is. Quite easily so.
The naive approach but still fast (“turbo”) and commonly used option is to use overlapping tiles and then blend the tiles using a gradient. E.g. a linear gradient or some other. It’s a very common solution for dealing with seams, and performant as well.
The only downside is that they’d need a couple of more tiles, which means more processing, so it’s a tad bit slower. The end result however is much improved since the seams become unnoticeable. A good trade-off in general though, in the quality vs speed department.
EDIT: Upon closer inspection of your image and given how tiling works, this might be a bit trickier for them to solve than the naive approach described above. The reason is because they likely don’t send global information to the respective tiling models, specifically some sort of color histogram. So if they were to just do blending, then you’d end up with a background that smoothly shifts to the darker or brighter variant that you’re seeing, and it’d end up creating sort of shaded islands of different brightness or color hue. That’s not what’s desired for flat or smoothly shaded areas.
So the approach for such would be to give each tiling model some (image) global information so it knows “how dark of a blue” the overall composition should be in a given region. The technical lingo for that is global optimization. It seems they’re just doing local optimization, which in layman’s terms means one tile has no idea what another tile is about, it just does its job in isolation with the tile it was given. When you put such tiles together, it doesn’t yield a very satisfying result as you’ve observed.
1 Like
Gigapixel AI is the only Topaz app being developed well and I still love it enough to resubscribe to its updates. As far a PAI and TVAI, their development cycle has gone the way of dog poo. I still use older versions of VEAI and TVAI to upscale videos, but I’ve given up on updates for now.
Am I the only one with Recovery quality mode producing black images? Want to figure out if it’s my hardware failing or if it’s a bug in the software.
In my experience (Mac) I have not seen this. Recovery is one of the most exciting new Topaz features right now! Do you see the black area in the preview, or only upon export? Please provide your computer specs.
1 Like
Thank you very much for the insight. Would love someone from the Topaz team to acknowledge the issue though. The uncompressed image is available above and the issue is easily reproducible. Hoping for the best as it’s not the only image I have that has this same issue!
My attempt to praise Gigapixel: As a hobbyist photographer, shooting animals and birds mostly in the wild, I use the Gigapixel as my primary or final tool to my satisfaction. The Recovery tool, even though it is still only in the Beta version, is a highlight. Recovery suffers from long processing times and I often have to simply cancel image editing when there is no result even after 20 hours. The computational complexity of the algorithm is in principle high and non-linear depending on the volume of data. It helps me to initially reduce the size of the photo (width x height, sometimes very significant). It’s true, my hardware is weaker, only a laptop with an NVIDIA GeForce GTX 1660 Ti, but so far, with a little patience, a nice result can usually be obtained. Of course, sooner or later some strange artifacts will be removed, for example, instead of a uniform magnified surface on a part of the duck, I see something that looks like a coarse fabric with the texture of large potato sacks – but it can only be seen at enormous magnification, so I can survive it. I’m just worried if improving the algorithm will make it unusable for “ordinary” people due to extreme hardware requirements.
The attached image (print-screen) shows a common dragonfly (a rewarding subject for a photographer 
). Cropped a 2517x1888 portion from a 6000x3368 photo, enhanced with Very Compressed 4x, then scaled down to 640x480 for the final Recovery Beta 6x to produce (for me) a pretty nice image.
So I hope that Topaz Gigapixel will continue on a favorable development path without TPAI imitations, which from my point of view have failed in a number of ways (much to my dismay).
1 Like
Completely agree and share your perspective. Didn’t renew TVAI because of that, and won’t until that team gets their act together. Gigapixel is currently the company’s star product. Usable, does what it advertises, and is shepherded by a great and engaging team.
3 Likes
Hey darkmatter!
I definitely reproduced the tiling that you were getting.
This a known bug in our Speed mode of Recovery that’s related to the color specifically in the pixels.
I’m going to bring this to the attention of the dev team to look at - thank you for sharing the image to test!
1 Like
aww thank you for the kind words!
