Others Saria Reclaimed Remake [v0.16.1.0] [UnfetterSangria]

3.00 star(s) 1 Vote
Dec 9, 2018
230
260
He can easily generate the art by using AI, which is freely available for everyone to use on CvitAI. And the art also looks exactly like the original ones. I don't know whats the problem here.

The model is . (Don't know if this is the one Hollowbankai is talking about) :

And the generated images, for example are:

View attachment 4048309 View attachment 4048314 View attachment 4048318
created by:
Among others:
- Wrong color of the shirt
- Mole is either missing or on the wrong breast.
- Light and shadows are off
- Arm in the third one just disappearing and the other with the wrong gauntlet
- eyes just dissolving and without the sharpness of the original
- only one of the pictures has the collar and not the right collar at that
- first picture with 2 elbow guards
- shin guards are wrong
- tits too big

Sure. They go in the direction of Vel. The dev could try to replace every picture with an AI product. Could redefine how the character looks in this remake and get rid of some of the pesky details, that the AI has a hard time to reproduce consistently.

But saying that they "exactly like the original ones" is an insult to Vel.
 
  • Like
Reactions: Humanbug

ConfusedCube

New Member
Apr 5, 2024
14
10
Among others:

- Wrong color of the shirt

- Mole is either missing or on the wrong breast.

- Light and shadows are off

- Arm in the third one just disappearing and the other with the wrong gauntlet

- eyes just dissolving and without the sharpness of the original

- only one of the pictures has the collar and not the right collar at that

- first picture with 2 elbow guards

- shin guards are wrong

- tits too big



Sure. They go in the direction of Vel. The dev could try to replace every picture with an AI product. Could redefine how the character looks in this remake and get rid of some of the pesky details, that the AI has a hard time to reproduce consistently.



But saying that they "exactly like the original ones" is an insult to Vel.

Didn't try to disrespect Vel or anything. All I am saying that these generated images are almost there.

Of course art created by people are more detailed, consistent, refined and polished. AI is not 100% all the way there yet. But it surely can get the job done.

Also, I doubt these images were generated for a particular goal or purpose, mainly for testing (or whatever other purpose he may have). But when generating for a particular purpose, we would write better prompts, use better checkpoint models, generate a much larger number of images and choose the best from them. Also, if a new lora model is trained, it may also generate better images.

A variety of techniques such as highres-fix, image2image inpaint, adetailers etc can also be used which will solve most of the problems mentioned here (namely, eye detail, shirt color, big tits, etc). And other details can be obtained by generating multiple images- trial error method.( And by multiple images I don't mean 100s of images. Decent images can be obtained after generating and refining the prompt 5-6 times).

And if all fails (mainly for smaller details or symmetry problems like mole on other side etc.) We have to edit those using photoshop or gimp or whatever - in which case it may need some extra labour (but is very unlikely).

Thus I was saying that the biggest problem in this situation shouldn't be such a huge problem after all.
 

Hollowbankai

Newbie
Sep 12, 2018
98
91
He can easily generate the art by using AI, which is freely available for everyone to use on CvitAI. And the art also looks exactly like the original ones. I don't know whats the problem here.

The model is . (Don't know if this is the one Hollowbankai is talking about) :

And the generated images, for example are:

View attachment 4048309 View attachment 4048314 View attachment 4048318
created by:
No this is not the one I was talking about I would have trained one from scratch. Also I would have not been able to train for XL because my GPU can't handle that, but XL would be the way to go because in base SD fingers still look like a nightmare.
 
  • Like
Reactions: ConfusedCube
3.00 star(s) 1 Vote