Boulders in nature photos often come out stringy, etc

In blowing up my nature oriented photos, Photo AI does a fantastic job on most plants, trees, water, etc. However, big boulders it often struggles with. I get various manner of unnatural looking artifacts that are often stringy looking or at times patchy/oil painting blobs looking, depending on settings. The main point being, all too often boulders do not up size well with a realistic look like everything else.

My guess is the machine learning needs more training on boulders/rocks or something? Everything else nature wise seems to blow up nicely so far on numerous photos.

Note that in the attached photo, I’m panning around on a boulder in the same general area. You can see the stringy aspects move to new areas with the panning and then blur is now there, the sort of blur I mentioned in my other thread.

2 Likes

Thanks for reaching out. Please send the following information so I can try to reproduce this.

  1. The original image file.
  2. The processed image.
  3. A screenshot or description of the settings panel so I can see what adjustments were turned on.

You can securely submit your image(s) to my Dropbox using the link below. Please be sure to send me a note to let me know you sent something.
Dropbox File Request

It looks like the Raw Remove Noise processing is removing too much detail from the image. Have you tried switching the model from Strong to Normal or vice versa? That may improve the processing so the details look less plastic.

Hi Lingyu,

Thank you for your response on this issue Lingyu. I’ll get you some files put together.

I was confused however that you mentioned the removing too much detail? That wasn’t the point of this post. The random blurriness was the subject of my other thread. The trouble in this thread, is nearly all of the boulders, rock faces, cliffs, etc., no matter how I process them, whatever models and settings in Photo AI, sections of the rock faces have what look like roots and branches growing across the rocks. I’ve tested numerous photos, and file formats. Boulder/rock cliff like textures, consistently, at least for me, are not enlarging in a natural manner. They look like “biological rocks.”

@devenpsc let me know once you have uploaded the files!

Files are uploaded! Thanks!

Thanks for uploading the files. I think the Remove Blur setting here might be causing the issue. Remove Blur is a weak sharpening which enhances details or textures.

In this case, the details on the rock are oversharpened to create the look that you are describing.

Can you try decreasing this setting and let me know if it’s working for you? You may want to decrease the Remove Blur setting and try one of the sharpening options instead for a cleaner sharpening result.

Thank you for your reply. I actually have tried every which way I could dream of on settings, for days, on dozens of photos. I was determined to figure it out, but failed. That’s when I joined the forum.

I tried remove blur on/off/part way, then trying other models/settings/combinations. Nothing helped the blurry patches. There’s some settings the stringy stuff isn’t as bad, but then the blow up in general looks terrible. I was using remove blur only because it was giving me the most natural results, at least for my purposes. The sharpen section settings often over sharpened or gave weird results, same with remove noise section.

Anyway, I just tried again on two of the files I sent you with remove blur both down and then off, while trying out low levels of the sharpen section also. Same results, blurry patches and stringy stuff on boulders, along with general results that don’t look good, unfortunately.

If you want screen shots of other efforts and settings let me know? Thank you!

1 Like

I see what you mean. I ran some tests and passed the images along to the researcher working on the upscaling models. He is investigating and checking if this is something we can improve.

I don’t have a solution at the moment for this behavior unfortunately.

Is this issue only in the preview or also in the output?

Both.

Consider reaching out to Kevin Shank on the forum. He’s been doing a deep dive on all of this and has some good theories on some of the issues.

It is absolutely coming from low-rez model with it getting worse with increasing scaling. It basically blurs all organic detail then tries to make smooth sharp continuous cartoon edges for what it thinks is important detail.

For some things that is exactly the low-rez result that you want. Suppose you have a classic window with a diamond pattern on it, it does a great job making them nice straight lines on a smooth window even if it was badly compressed and noisy, whereas standard makes it very organic mess depending how much you play with deblur/denoise/dejpg. So it is very much needed as an alternate AI

The problem is with nature photos it is going to hallicunate and connect the dots (which are actually pixelated natural textures) into these strings in a awful attempt at being cartoon sharp edges smooth elsewhere. Thin streaky white things that look like white grass on the rock but obviously that makes no sense not even lichen look like that, cartoon black rock crack that sticks out badly, streaky grass shadows with no texture when that is actually organic streaks in the rock pattern itself, log is completely posterized.

This is a fair use transform of a photo copyright from devenpsc, I leave it to them to provide source.

low-rez
0 supress noise (aka denoise)
100 remove blur, (aka deblur)
no remove compression option (aka dejpg)

standard
0 supress noise (aka denoise)
100 remove blur, (aka deblur)
100 remove compress (aka dejpg)

the only strings is grass, the jpeg mess is cleaned up for smooth edges, but the rock and wood are proper rock and wood textures. The only strings are arguably grass and wood cracks, maybe a few near rock edge is confused by jpeg watercolor bleed noise or maybe that is just fine grass/weeds

Side discussion but devenspc also has a bug for patchy blur/details and I think it conspires with these different ai goals. depth buffer blur also behaves differently between the two ai

deblur pushed up on standard is pushing the blur out to only the depth confusion where rocks/grass meet, lessen deblur it starts blurring more into the rock extending the depth mask from the grass. It is actually getting the depth mask wrong as it has no concept of front or back grass as I think it relies on bright/detailed dim/blurred to clue the depth mask and not really true depth plan object sorting.

low-rez wants to bring out all those white highlites with streaky connected hallucinated as a cartoon foreground, and smooth the gray parts of the rock into a blurry background - classic cartoon tricks to pop 3D out of 2D so that conspires with the depth mask. those strings on rock must be front, that gray blur must be back.

I think neither is really fixable without the intended target of each AI being degraded.

You do not want to fix low-rez to let the dimmer details in and not connect the dots to prevent the posterized cartoon because you need that to make a patterned window panes work

Interestingly auto got worse in 1.3.2 standard was more prefferred, whereas 1.3.4 low-rez was more preferred and I think it is because 1.3.4 has moved more to wanting blurryness to avoid artifacts but low-rez likes the blur as it makes it easier to connect the dots.

Now you might just say why not just let user do standard all or low-rez all knowing the image content since auto seems to have a problem (and even then it cannot know what is artistic intent - maybe you want that cell-shaded posterized look on nature) The problem is you cannot do that in photo as it has copy settings to all , when all you want to do is set the AI model on selected images and let it pick the model parameters - and no way will you do that manually on a large batch image set. So you are reliant on AI to pick the model. You can do that with gigapixel because of its spreadsheet UI option (and I just posted an enhancement request for photo to have that)

I think devenpsc said in an older version it varied with preview zoom/pan but it got fixed (probably just was running AI on visible part) but it seems to be current version save is consistent with preview (probably fixed by running AI on even hidden part of image). gigapixel might still have that problem in varied preview as it is not updated since winter.

I do have a seperate issue that with pretiled seamless textures it is not coming to the same exact solution in the repeated parts - probably has to do with model size coverages varying as the image is scanned they do not line up exactly - but that can mislead into thinking it is the preview problem but I do not think it is.

Hi everyone – Lingyu is out of office for the next few days but we’ll resume looking at this once he gets back!

Normal, Mormal V2 Beta, Strong and Strong V2 Beta all have the same problem for other than boulders that have textures, but Strong not so much because the detail recovery is poor.

So regardless of which sharpen algo is used, they seem all similar issue connecting the dots. You see it works well for the kitty hair sharpening that though it looks unnatural even there as it strings into the rock.

Seems there is no AI that is leveraging the pattern recognition in AI to segment textures this is hair or feathers or grass vs. rocks or dirt or flesh and it is just a conventional image processing sharpening.

So it seems to need multiple subject segmentation that goes beyond the existing fore/aft subject control for sharpness regardless of AI or enhance sharpening. It will probably need manual editing as well.

You are right, in fact I don’t use DeNoise or Photo AI on images like that as there are better options out there that don’t destroy texture, don’t make detail recovery a guess and do differentiate between hair and feathers.

Don’t know how we got to this point because I believe the AI actually flattens the image before processing and, as you said, the sharpening seems to be an insurmountable issue without manual intervention.

Forgot to add the unprocessed image here:

I’m curious what programs you are talking about. I’ve done a deep search for every program I could find that does enlarging of any type, and none of them do what PAI/Giga can do, that I’ve found. If you know of one, please share? All the ones I’ve tried, including PS super resolution, etc., do just plain old enlarging, or a super crappy job with their AI, far worse than Topaz does. Not to mention the various competition’s rendering times are absolutely ridiculous.

The discussion is about problems with noise reduction. If I enlarge the image I just enlarge the problem.

But for resizing there are a couple of new options out there that don’t destroy the details but GigaPixel is still acceptable if you can get a noise reduction process that doesn’t destroy the details.

Understood on the noise reduction. So, what are the new options you have heard of for enlarging?

This topic was automatically closed 7 days after the last reply. New replies are no longer allowed.