SHP Forums

Oh! Nice. I was able to take that, resize it a little, layer it and blend it to my smudge colorings - and came up with this.

Your scan really helps because you can clearly see her fingers and such, and way more features to her face.

I noticed a couple of things, so I tried my hand at it.

Nice! I like the eye to add the 3-D Animated Adventure bit at the bottom too!

We may yet be able to replicate a pretty damn nice version of the cover.

You can see a trace of the 3-D Animated Adventure in MI's reconstruction, even if it is not more than a smudge.

I took on Collector's latest - cleaned up how I could see lines on the left of the image and cloned to remove those lines. Cleaned up a few of the white lines too, to make a less damaged box.

If we could get a model specifically trained for these fantasy type of pseudo-realistic paintings, we might be able to do better. Perhaps with training a model on high resolution scans of the other KQ box art that we have? May be worth a try. Stylistically they're all quite similar. I've never tried training a model before. Might be a good time.

MusicallyInspired wrote: ↑Tue Jul 27, 2021 3:17 pm If we could get a model specifically trained for these fantasy type of pseudo-realistic paintings, we might be able to do better. Perhaps with training a model on high resolution scans of the other KQ box art that we have? May be worth a try. Stylistically they're all quite similar. I've never tried training a model before. Might be a good time.

I assume you mean a digital model... when you say "train a model."

Messed around with darkening/shading some of um, the haggy lady's dress, darkening the storms, lightning around the light... (How am I forgetting Edgar's "mom's" name right now?)
Loletta?

I am not qualified to train a physical one.

And her name is Lolotte.

MusicallyInspired wrote: ↑Tue Jul 27, 2021 4:28 pm I am not qualified to train a physical one.

Figured that might be the case. I've never messed with a digital model?

You mean like HEROFORGE and stuff?

ESRGAN is an AI neural network that does all this fancy new quality upscaling. To get it done you need a dataset model that is trained on certain types of images to get the upscaling you want. There are models trained for anime and manga, models trained for cartoons, photorealism, JPG compression artifacts, noise, blur, you name it. You have to use the appropriate models for whatever you're trying to upscale/restore or you will have very different results (ie- if you feed a model trained on cartoon images a photorealistic image it's not going to come out great).

And you can make your own dataset models if you feed the trainer two sets of images; one with full resolution images (what you want the end results to look like) and one with compromised resolutions or quality, depending on what you're after. The more images you feed a model during the training process, the smarter it gets and will have more favourable quality when used for upscaling/restoring. It's very similar to all the deepfaking that's been going on on the internet.

Think of it like ESRGAN is a printer and the dataset models are the paper. If you print on regular paper but tell the printer to expect premium quality glossy paper the printing results aren't going to go that great.

But yeah, it has nothing to do with miniatures of any kind, digital or physical.

Or model like what they used to call "mannequin" (a real person).

Messed around with this some more - funny result.

Used Rosella's face from the normal KQ4 cover to mimic the face and Lolotte's eye.

Um, it doesn't look all that different?

*zooms in - and in - and in* Ah! Now I see it.

Was bored (45GB file uploading, and movie making in another) - so I came back to this and thought, "Why not super impose real faces and edit them?"