KoW - D.R.E.A.M.

[MrFlibble]

Nulfaga with the yellow teeth, evil half insane grin, tangled hair. The blind oracle.....

Yup, these look pretty good, but I cannot but notice that the AI has some trouble making the eyes look straight.. they all seem to suffer from strabismus to greater or lesser extent. But it likely won't be as noticeable in-game I suppose.

All images in the prev post are made with the low AI input (exactly as the "boring" dark skinned lady on the left here) but when u look closely to the prev post, they are really good. I will just say that But it takes time... and lot of processing power.

Everybody can make the images on the right look good, because its basically all done by AI..... The less AI there is, the more u need to know, to make it look good.

Thank you for posting this example! I can now clearly see that you have more or less found the right balance to improve the image without letting the AI deviate from the source material too far. It's a pity though it seems to neglect the feet so much, while greater AI input still produced wrong legs for the lady in most images... it still has a lot to learn.

I also find it interesting that the AI went to reshape the body a bit in the more heavily modified versions, I think it kind of felt the proportions are not exactly right because the original sprite was likely drawn for a 4:3 display with non-square pixels, while your upscale is based on the "raw" sprite with 1:1 pixel ratio.

Still, please feel free to post any interesting results that you might get from the AI, even if they're not exactly what you're aiming at!

[King of Worms]

Hi, thanks for a feedback

Face is the most difficult part of human body for AI to produce, because our brains developed for millions of years, to read the faces. So we detect even a miniscule details and are super sensitive to it The faces are quite improved in the 2nd batch of images tho, eyes too. We can always find minor issues, but when u compare it to the sprites its based on, than u realize... that those had 10x more issues And yes, eyes are the pinnacle... its literally few wrong pixels and it breaks down.

Face and hair is what makes the NPC. Legs, not so much. Thats like 10% of the overall impression in my experience. Esp in DFU, when u come close to NPC, all u see is face and upper body. You can see legs, only from distance, or when u crouch to examine em up close. When u stand close to NPC, and look down, the legs are very small

Well I think it changes the shape of the body when u increase the AI input, because the model was made in a way, that with the prompt it has (text describing the image) it tends to imagine the body in a certain way. So if I changed the age in the prompt, it could made the person more fat etc... Other models would produce different results. This model likes to produce perfect versions of ppl, when u let it unhinged.

Anyway, heres some more. Redid the princess, as she was way off in the 1st example. I increased the AI input by 1%. So the difference between the faces and the bodies is not that big. Faces are generated by different AI stable diffusion than body... thats the only way to make the faces look this good. Its the model which only does faces.

The max detail I can get on my 10gb VRAM card is roughly 2 200 000 pixels. If I get just a bit over that, the rendering time increases 10 times. If I had 16gb nVidia card, it would be better. But with the price of such cards, one can be glad he has something...

PS: made some tests with working on the original Daggerfall art as a source and its impossible. Im quite glad for that. Because to get the results I present here, u simply need to put in the work in making the original sprites high res AKA my art I created is necessary for this. I hand drawn a lot of shadowing and details, faces etc, and this all gets taken in account when producing these high detail versions

test 3.jpg (1.12 MiB) Viewed 1194 times

[MrFlibble]

Anyway, heres some more. Redid the princess, as she was way off in the 1st example. I increased the AI input by 1%. So the difference between the faces and the bodies is not that big. Faces are generated by different AI stable diffusion than body... thats the only way to make the faces look this good. Its the model which only does faces.

These all look very good! The realistic faces reminded me of NPCs from The Quest somewhat.

I still wonder what the Daggerfall flats would look like if a video game artist took to redrawing them manually at this resolution.

PS: made some tests with working on the original Daggerfall art as a source and its impossible. Im quite glad for that. Because to get the results I present here, u simply need to put in the work in making the original sprites high res AKA my art I created is necessary for this. I hand drawn a lot of shadowing and details, faces etc, and this all gets taken in account when producing these high detail versions

I know your sprites are already in the mod, but perhaps you could post a few side-by-side comparisons between your source images and the results that you get?

[King of Worms]

Damn, the Quest has its charm! Looks like those old RPG crawlers

I can post some comparisons indeed, will do so.

Just came here to post the updated Court screen. This one took my GFX card to its knees, but I managed to render 4 images in 1920x1200 somehow and than combine it to this one.

I like how "room filled with smoke" prompt made the smoke lie on the ground.. I think this is a very true to the original recreation.

CORT01I0.IMG.png (1.86 MiB) Viewed 1070 times

[King of Worms]

And as comparison I will use this image. Because I was totally not able to make the left image any better, as the face is in such position u need to really know how to draw faces, to pull that off.

And the result completely nailed it. I always render many variants, and than pick the one which has most similar face to how I imagined the NPC all the time.

If you zoom in on the new version, and check the details, its so nice. Not just the face, golden accessories, but also the leg, knee, the golden shoe. Its just a perfection in my eyes

Compare.jpg (607.22 KiB) Viewed 1069 times

[jayhova]

The problem with AI is one of anachronistic context. This is a fantasy game set in a medieval world. You get these models that that look like fashion models with modern makeup wearing designer clothes made of modern fabrics. You also very quickly loose the artistic feel of the original for what is AI realism.

[King of Worms]

The problem with AI is one of anachronistic context. This is a fantasy game set in a medieval world. You get these models that that look like fashion models with modern makeup wearing designer clothes made of modern fabrics. You also very quickly loose the artistic feel of the original for what is AI realism.

I agree but thats when you overdo it. But this is not the case here in the latest iterations of my NPCs. I take good care of preserving the original colors, proportions and low detail. It was demonstrated in the image with 8 same figures of a dark skinned lady. As u can see in the last image on this page, the clothing and style stays the same. Also, in the 1st image on this page... This is because the AI input is set to very low, and the colors are transferred from the original image to the new image I generate a lot of variants, and than select the one which has the face resembling the original the most.

I for sure avoid creating these flashy model like figures with garments from a fashion house. Completely agreed that that approach would never fit Daggefall. Its just what most ppl go for, as it looks shiny and nice, at a 1st glance. I aim at something very different than that.

This game is 90s sword and sorcery fantasy style. Theres a lot of nice half naked females, strong men... its something I grew up on, so I believe I can pull this off. I know, how this style is supposed to be. And I will do my best, to achieve that vision, using the best tech I can lay my hands on.

[King of Worms]

I opened DREAM & DFU modding Discord server, everyone is welcomed

Invite link
https://discord.gg/Pxsg3B5A

[Taryl]

I agree that you are onto something. When you get it locked down, I think it will do what you want.

At the moment in the images you showed there is a noticeable difference in the resolution of the faces and the rest of the model that distracts from the quality of the work. For instance, the first Elysana you showed has more software errors on the model, the crown and hand are misshaped, for instance, but it looked like it was all one piece. The second version was more correct and higher 'quality', but the face did not look as if it belonged with the rest of the model. All of the faces in that row gave the same impression, though the boy's was least jarring. It's true that the face is most important, but it looks as if the face is 4k and the rest is 720p. I also would have thought from the original image that Medora Direnni was wearing sandals with a gold buckle and straps, but it seems the software disagreed.

I hope that doesn't sound disrespectful because I have always liked DREAM and been impressed by your efforts. The new image software is what I'm less impressed with, but I guess you will tame that beast.

[King of Worms]

I think its best to wait how it will look ingame when/if its finished .)