Man, that’s just messed up. It shouldn’t be so difficult to resolve that you’d still be waiting after all this time since it’s obviously about some kind of check during install that comes back with the wrong result.
How do I access those profiles? I couldn’t find anywhere to set the grayscale profile as I load greyscale photos for upscaling usually.
I think it just means that GP is able to preserve the grayscale profile from the input image to the output image.
2 Likes
alanhusband-677420 is correct =)
Models are conflating colors for different image regions.
I have the following source image:
It was then downscaled to 256 pixels wide, and saved using the traditional 75% jpeg quality (too high for any professional use, but typical for internet images).
Then the image was upscaled using the three usable models (High-def, Low and Standard).
High fidelity
Low res
Standard
Observation
- As can be seen with all of them, the color of the nose is for some reason applied to both eyes for High and Low, and to the left eye for Standard.
- Additionally, the color of the nose, which is plainly discernible even in the downscaled image, is turned into gray fur color and texture for the High and Standard model outputs. For the Low output the model uses a brownish color and decides to also color the entire mouth contour with that same color.
- The original nose color also bleeds into all kinds of places on the fur.
- The problem with red-ish color can be seen reflected in the pillow ass well, where the models exhibit clear problems recalling (paying attention to) the actual colors in the source image, and instead hallucinates some other color (bright red in this case) for the output.
(This makes me think of VIT transformer models, where if an encoder has problem with its attention structure, produces attention masks that throws the decoder off. In this case, something like “hand-wavy, yeah, there was some brown in this vague vicinity. When decoding, try your best to fill in the blanks for that lousy description I just gave you” and then hands off that unhelpful bit of information to the decoder which has to work off of that to construct an output image).
So there’s definitely something structurally wrong with all the models when it comes to working with warmer tones. Hopefully some of the researchers can pin-point what exactly is causing the models to get so consistently confused, and then release new versions of the models that exhibit less of color region confusion / hallucinations.
4 Likes
Hey guys,
Is there a way to fix these weird rectangular artifacts when trying to improve this photo in the Recovery (Speed) mode? These are present only in Speed mode, which I’m most interested in.
Here’s the original picture:
Here’s the recovered version:
Thanks!
I’ve found Recovery to be quite bad with empty background areas in general - it’s always looking for detail to enhance so will create some pretty strange artefacts. Where possible, I’ll upscale again with a different model and copy-and-paste the undamaged bits in.
On Win 11 update from 7.0.0 to 7.2.1 keeps crashing.
On Mac Sonoma update working
Well, you see, it’s still not fixed. It’s the same thing with Photo AI. I’ve had the problem since Gigapixel 7.0.1 and Photo AI 3.0.3. Before that, no problem. However, I have to use the command prompt to run it as administrator, and it works.
@yuan.liang can you shed some insight on this?
Thank you for sharing!
With speed mode, the output quality compared to Quality is as expected where outputs are not as ideal.
As far as weird rectangular artifacts, I can certainly speak with @yuan.liang to find out more about what’s happening here.
Hey Olaf!
Thank you for letting us know!
Could you submit a ticket with our Support team and we can communicate with you more directly to review?
You can include Windows Installer logs and we can check those out.
With Recovery being a little more generative I can see why the Ai model searches for something to recover but I’d be really curious to see your results.
Hi Harald,
Sorry to hear that this is still not fixed.
Have you already started a support ticket with us where we can chat directly about the issue?
Hi,
Thanks for investigating the models! I am not fully sure, but the reason can be as follow. Let us know what is your thought.
I attached a screenshot switching between 1) high resolution source image, and 2) stretched input image (not applied any model). You can see there is color shift (dimmer for nose and pillow in input image). The color shift is caused by, when downsizing image, the high frequency colors (bright colors in this case) is averaged out by surrounding darker pixels.
What the models’ task is, by seeing only the input image, predicting the original image. We can see the models are aware that, the nose and pillow pixels should be brightened up a bit when upscaling. The models know this because they have observed this effect (high frequency color gets washed out) during training. However, they do not exactly know how bright they should be. They just try and give different results.
I think this explains they turn the nose/pillow into different colors from the ground-truth (high resolution image). For example, for pillow pixels, all models lights up the pixels - hifi model lights up more, low reso model lights up less. However, the reason why they cannot give exactly the correct color as in the ground-truth, is that all models are not perfect, just like they cannot make good fur texture as in original ground-truth. But please rest assured, we are keeping to improve this all the time.
3 Likes
For this, @xuan.liu is the better person to help answer.
1 Like
Hi there.
Yes I had already made a ticket some time ago. The only temporary workaround they found (and it works) is to do as command prompt as administrator. Type “cd followed by the link to the location of the installation file” and validate. Then type this command with the file version number: “msiexec /i TopazGigapixelAI-7.2.1.msi” without the quotes.
Example of a screenshot I took when releasing version 7.1.2
Hi Esther,
got it resolved. It worked doing three updates:
a) from v7.0.0 to v7.0.5.
b) from v7.0.5 to v7.1.2
c) from v7.1.2 to 7.2.1
this way 7.2.1 in the end installed successfully while trying
to leap directly from 7.0.0 to 7.2.1 did not work
2 Likes
thank you for the update!! glad Gigapixel is working for you
1 Like
Hi Esther
Unfortunately, I’ve deleted most of the best examples now, having written them off as a lost cause. The following illustrates the kind of results I was seeing though.
Original image:
Output with Recovery:
Note the top right corner in particular; Recovery seems to have interpreted the small amount of compression there as some kind of fabric pattern that it needs to enhance.
3 Likes
