Gigapixel v8.0.3

[Lingyu]

Hello!

This patch for Gigapixel removes the 12,000 pixel limit for Redefine and some other bug fixes. We want to get this out so anyone with 8.0.0 can benefit before 8.1.0 releases.

Please continue to share your images and tests and let us know of any issues you encounter. We’ll be updating Gigapixel regularly to address your feedback.

Thank you!

---

v8.0.3
Released December 4th, 2024
Mac: Download
Windows: Download

---

Changelog:

• Removed 12,000 pixel limit on Redefine
• Fixed visual inconsistency in progress bar when processing
• Fixed old model file deletion not working on Windows

---

Lingyu Kong
Technical Product Manager
Image AI

[Imo]

Great news Lingyu.

[Harald.De.Luca]

So here are some really cool results.
This is the original enhanced without scaling with Redefine at the following values:
Creativity at 4
Textures at 4

From this new image I upscaled to 8460x6342
The Redefine settings were:
Creativity at 2 and Textures at 3.

Of course I did an automatic prompt with the Dzine.AI site for speed.
The prompt is as follows:

“A red panda walks on a horizontal ladder made of tree trunks, its fur contrasting with the earthy tones of the surrounding wood and forest floor. The panda’s tail sways gently as it moves, adding a dynamic element to the scene. The background is filled with lush greenery such as bamboo leaves, providing a natural habitat for this charming creature.”

I then made several variations with stronger or weaker settings, starting either from the basic image or from an already enhanced version.
There’s also a passage with the Recover model for certain small details.
Then you’ll notice that on all the variants below, there are some elements that are good and some that are weird, and also there are some tiny red pandas that have generated themselves on areas that shouldn’t hahaha.

But as I said earlier, each time I took the correct elements from each image to recompose a final one.

So here, I’m redoing the wooden platform with some of the decor that could be salvaged. I’ve refocused the original image with TPAI’s SuperFocus model.
I also upscaled x2 at the same time.

Well, as you can see, the animal is blurred, but debruited. But in this case, I don’t really care, because in this case I’m recovering the foreground scenery, since I’ve already reworked the details of the animal.
So next, I just upscaled to 8064x6048 with Redefine from the variant with the superfocus.

So, I mainly recovered the foreground part + part of the background. I kept the best part of the background on the left from a combination of the previous best variants.

To conclude, I think this is the final version of the image.
If, however, I don’t add a blur in the background to bring the animal into sharper focus

Tell me in the comments what you thought of this new approach to image enhancement/restoration.

Just to compare with the 2 pictures side by side

[Imo]

But the old bug where the app won’t process Redefine after making changes to the settings persists.

[Lingyu]

Could you direct message me with logs and a screen recording if possible? We haven’t been able to reproduce this yet, I’d like to fix it of course.

[unrelent]

Interesting and very informative for me. I personally prefer a somewhat blurry background, because it primarily attracts me to the main subject of the image, but it depends on the intention and goal of the editing.

It seems to me that the red panda’s face is shifted a little towards the dog-like type of face; maybe the result of Redefine? But maybe it’s just my imagination. Anyway, thanks for the inspirational description and examples.

Higher Texture (3-5) with Creativity 2-3 works well for me on blurry photos of birds from a distance (after cropping), but it also creates a lot (sometimes dozens) of artifacts – which then have to be removed manually with something with the remove object function. On the other hand, it creates nice feathers and other details. Redefine gives me here better results than Recover.

A cutout from the original image (my camera, moreover handheld, can’t do much more than that from a distance of about half a mile):

Result of Redefine (4x, Creativity 3, Texture 4, I couldn’t find a better combination). There’s a whole flock of strange bird-artifacts around the heron, invented from ordinary clover in a field – the AI ​​is sometimes very inventive:

Well, here it is with the artifacts removed. Not a miracle, but better than a white spot on a gray-green background. The clover doesn’t fit there, but the heron chose the field where it found food, and its goal was not to make a mannequin, as I had hoped:

[gary.5bellies (Gary Loader)]

Redefine still only works on cloud-rendered images. Asking for any preview crashes back to the desktop as a standalone or plugin. IPC is 7950x, with 4090RTX and 64GB ram.
Please advise
Gary

[Harald.De.Luca]

Well, the final image is rather nice .

[Harald.De.Luca]

And for my red panda, the muzzle is indeed that of a red panda.

[JYG]

Only because you asked for comments:
Don’t get me wrong, but when I compare the original with your ‘final version of the image’
I honestly don’t see much improvement justifying some of the effort.

• The background looks fine to me in the ‘original version’ and although it looks different it hasn’t improved all that much, maybe a little. I’m still kind of debating with myself on the look of the background really. I get where you’re going for with the background, but the final version to me isn’t overall ‘nicer’ to look at, it just looks a little ‘different’.
  The leaves look less detailed and less distracting but they also turned out a little dull and bland at the same time, and I think I like the slightly darker green tone of the original leaves a bit more as well.
  What I also just noticed as I write this reply is how the look of the plants in the background varies from above the bridge to below the bridge. (This is of course something that should be quick and easy to fix so the plants in the background look uniform all across the image.)

• The actual problem that I personally have with the ‘final version’ though, is with the altered wood textures on the tree logs of the bridge.
  To me they just look unnatural in the ‘final version’ and look much better in the ‘original version’ (as well as in some of the other intermediate image results).
  I just think the wood textures on these logs were "broken by AI’ making the original and some of the intermediate steps look significantly better than the end result does, in my humble opinion.

The ‘broken’ log textures just bother me personally far more than the (minor) changes of the background and of the panda help the improve the image.

[michaelezra]

There is an inconsistency in Recovery in the residual amount of noise, depending on the size of the input image / area being developed in the Preview.

Illustrations are zoomed in

Preview area: Small

Preview area: Medium

Original image:

[plugsnpixels]

I keep saying here that Redefine is for REDEFINING, not necessarily straight enhancing. I use it pushed to max Creativity to get some incredible results out of abstract digital mush, using Gigapixel as an AI art app and having tons of fun doing it.

Since Redefine is doing its “redefine” thing at even low levels, you might as well run with it. Use the “classic” models for normal enhancement (they are excellent), along with Face Recovery which is also great for genealogical work (bring great-grandma back to life!).

This reply now gives me an excuse to post more soothing “happy place” artsy examples from Redefine, detail crops from much larger 6X renders… I intend to repurpose these types of results as “art” in various forms (textures, “paintings”, etc.).

A sample artsy workflow (WOMBO > Gigapixel Redefine > artsy app or image-to-image):

Tight crop:

Add oil paint art effects:

More artsy!

Use the original result as inspiration for image-to-image in DiffusionBee:

Another before and after:

100% crop:

Image-to-image:

I am now NormAIn Rockwell! I may never leave my desk again!

PS: I was curious what would happen with some random abstract painting from the web using Redefine with high Creativity, and here you go (Cloud render at 6X). The WOMBOs at least have some semblance of subject so that seems to help make the results more interesting.

[lhkjacky (Jacky)]

@Harald.De.Luca

The background of the image and the bamboo leaves are really well recovered .
The quality of the wood on the platform and the tree on the far left are less well recovered, however. As for the animal, it’s nice apart from the clawed paw tips. There are a few odd and unnatural features. But given the quality of the original image, the overall recovery is already nice .

Thank you for the advice!
I used ComfyUI inpainting to correct the claw, and I think it has improved slightly compared to my last result.

Here is my final result:

Here is an animated version.

Imgsli

@plugsnpixels

I keep saying here that Redefine is for REDEFINING, not necessarily straight enhancing.

Yes, I agree as well. Even when I use the lowest creativity level, Redefine AI tends to alter the image more than I prefer.

[plugsnpixels]

Yes, and the “Creativity” label is a clue as well

[jamie_wood]

They hurt my eyes.

[kevang]

G8 on the left, original image on the right. This is after upgrade to 8.03 … this is just getting worse. This took 31 minutes to generate (original image is 2048 x 1152, not a large image by any stretch)
So far i’m not seeing anything of value here. If the software isn’t optimized for Mac, then why release it?

[Harald.De.Luca]

Yeah, things are looking up. But the front paw is weird, I think. Too long. And normally it should have 5 claws for its 5 fingers.
Here’s an example from the Internet

[Harald.De.Luca]

Nice video. What AI did you use?
Well, at the end of the video with the panda moving around, I can see that the added scenery is quite different from the rest (mainly the wooden platform). The rest, with the leaves, looks good. The panda’s movements look realistic.

[Harald.De.Luca]

In fact, it has 6 fingers and not 5 as I thought.

[10061996]

This. I tried Redefine on AI generated photorealistic animal images and I have yet to find a setting that does not grossly exaggerate fur texture. The idea of “enlarging” by filling the additional pixels with AI and implementing that into GP certainly has a bright future to it but for the moment, there are even free models such as 4x_NMKD-Siax_200k or SUPIR that do a better job, at least for the stuff I generate.