They were modern, elegant and powerful. Humans highly modified. They were immortal, and for them the 100% organic human was not better than an insect.
Eran modernos, elegantes y poderosos. Humanos muy modificados. Eran inmortales, y para ellos el humano 100% orgánico no era mejor que un insecto.
AI Art Experiences
After a little break, and some days of experimenting, I'm back. After have seen some AI portraits from @dbddv01, I was eager to give it a try. By the way, you should take a look to his blog, it's amazing how he is building an ecosystem with his creations. So he inspired me to make some portraits, and it has been very challenging.
Portraits are not the strongest point of this AI, that's for sure. Sometimes I got well done faces by luck, often because I have add to the prompt an artist who painted many of them. But ask for a portrait in particular, doesn't work most of the times. And anyway, the result is a mess. The AI suffers to place all the elements of a face, the proportions... so I wanted to try @dbddv01 approach, and use an initial image to help the AI. Do that improved a lot the result, but continued having some problems, and the more iterations the more increased. So I have been doing lots of runs along some days and just observing what happens. How the AI interprets the image given. And also some research in the internet.
The AI analyzes the image, and split it in areas. It considers the lines that are draw. Also, each pixel is checked, their values of lighting and color are extracted, and areas created around these pixel values. One example can be the difference of the brightness between a highlight in metal ball and the nearby areas that are not getting directly the rays if light. The AI gonna consider it to different areas and create a shape that is not gonna be displayed but considered for future use when painting the image. Same with the colors, any different patch of color gonna be considered as an area. And used when painting as well. So considering all this, what I have done is to edit the start image in photoshop, to avoid unwanted results most as possible.
An example can be how I fixed a recurrent problem that I had with the lips. Lips got giant and very distorted, split in pieces, in every run. The split was happening due the highlights on them in the start image. It created some kind of lines in the lips and the AI considered them the borders of different areas. Had to think a lot about why the lip was getting giant. It seemed the AI was merging the lip and its shadow. But the shadow seemed black, the lip was red... I was missing something. At end I remembered that the shadows still having color... so, used Photoshop to increase the brightness of the start image, and yes, the shadow was tinted with the same red than the lips, so the AI considered it a whole area.
So to solve all this stuff what I have done is, desaturate all the start images, and use the levels to increase the contrast and decrease the dynamic range most as possible. And also delimitate areas using the clone stamp, the brush, and any tool to make clear to AI how should split the image. By the way, the start images have been made with Artbreeder.
Changing the subject, there are not new updates for the notebooks, and in fact I think gonna downgrade them. The anti-disconnect features are not working anymore. Google have been making some updates, to stop the insane amount of bots using Colab. Also seems they did something else, and the anti-disconnect requests to the server gets rejected. So I'm gonna remove those from the notebooks eventually.
And that's all, hope you like these AI portraits. Regards!
Después de un descanso, y unos cuantos días experimentando, he vuelto por aquí. Después de ver algunos retratos hechos con IA por @dbddv01, me entraron muchas ganas de probar. Por cierto, deberíais echarle un vistazo a su blog, es increíble como crea un ecosistema con sus creaciones. Así que me ha inspirado a intentar hacer unos retratos, y ha sido todo un desafío.
Los retratos no son lo mejor de esta IA, os lo aseguro. A veces me salieron algunas caras bien hechas por pura suerte, a menudo porqué añadí al prompt un artista que había pintado muchos de ellos. Pero pedir específicamente un retrato, no funciona la mayoría de veces. Y si funciona, igualmente el resultado es un desastre. La IA sufre para emplazar correctamente los elementos de la cara, las proporciones.. así que decidí probar la táctica de @dbddv01, y usar una imagen de inicio para ayudar a la IA. Eso mejoró mucho el resultado, aunque seguía teniendo algunos problemas, los cuales se incrementaban a cuantas más iteraciones. Así que me dediqué a lanzar la IA muchas veces a lo largo de unos cuentos días y observar que sucede. Cómo la IA interpreta la imagen que se le ha dado. Y también un poco de investigación en internet.
La IA analiza la imagen, y la separa en áreas. Toma en cuenta los trazos. Y también cada pixel es analizado, sus valores de luminosidad y color son extraídos, y en torno a esos valores crea las áreas. Por ejemplo, imaginemos los brillos en una bola de metal, la diferencia entre sus brillos y las partes de al lado que no están recibiendo la luz directamente. La IA considerara esas zonas como areas distintas para usar esa información cuando pinte. Lo mismo pasa con los colores, cualquier zona con un color en particular se considerará una área. Y se usará también esa información, a la hora de pintar. Así que tomando en cuenta todo esto, lo que he hecho es editar la imagen de inicio, para evitar todo lo posible resultados indeseados.
Un ejemplo de esto puede ser como corregí un problema muy recurrente con los labios. Se volvían gigantes, deformes, separados en varias partes, en cada intento. Lo de que fueran separados en trozos era debido a los brillos en los labios. Creaban una especie de líneas que le hacían pensar a la IA, que eran líneas que delimitan distintas partes. Entender porqué el labio se volvía gigante me costo mucho. Parecía que la IA estaba tomando el labio con su sombra, como si fueran una sola área. Pero la sombra parecía negra y el labio rojo... algo se me escapaba. Al final me acordé de que las sombras también tiene color... así que con Photoshop, le subí el brillo a la imagen, y si, la sombra tenía el mismo rojo que los labios, así que la IA lo estaba tomando como una sola área.
Para solucionar todo esto, lo que he hecho es desaturar todas las imágenes de inicio, subir el contraste y acortar todo lo posible su rango dinámico. También delimitar las áreas con el tampón de clonar, el pincel y cualquier herramienta para dejarlo claro a la IA como debería separar en areas la imagen. Por cierto, las imágenes de inicio han sido hechas con Artbreeder.
Cambiando de tema, no hay nuevas actualizaciones para los notebooks, en realidad los voy a degradar. Lo sistemas anti desconexión ya no funcionan. Google hizo algunas actualizaciones, para frenar el numero malsano de bots usando Colab. Parece que hicieron algo más, y las peticiones que el sistema anti desconexión le hace al servidor son rechazadas. Así que las quitaré de los notebooks en algún momento.
Y esto es todo por hoy, espero que os gusten los retratos hechos con la IA. ¡Saludos!
Notebooks
Enhanced Original VQGAN+CLIP Notebook
Mse regularized Modified VQGAN+CLIP Notebook
VQGAN+CLIP with Video Features Notebook
AI Artworks
All the content of this post is from my own.
Todo el contenido de este post es de mi autoría.
Images generated with VQGAN+CLIP, upscaled with Real-ESRGAN and edited with Photoshop.
Imágenes generadas VQGAN+CLIP, escaladas con Real-ESRGAN y editadas con Photoshop.
100% AI free writing.
Textos 100% libres de IA.
@jotakrevs, thanks for the mentions. And congratulations for the work spent on this challenging issue with faces. I will have to study your findings... But anyway, it looks like you greatly succeed to take back control on the AI. Those faces are just Wow... greatly balanced !
Thanks for your nice comment! Well, I'm happy with the result of my experiments, but the faces still far away from be perfect. Even delimiting all the areas in the start image. The noses and ears are just impossible to control... plus random stuff, at end the AI can decide ignore the shape of an area and expand and transform it as wishes. Next thing I want to try, is use just line art, without any kind of shading, and see what happens. Hope my findings gonna be useful for you.
Regards and don't forget your !PIZZA!
Dude... these are all badass! I had to take some time off from playing around on the AI but i am getting back to it. I still haven't posted on it... so i better get to work! hahahhaa Thanks again for all your contributions bro!
Yeah dude, go to work and deliver us something cool! :D
Thanks for your always kind comments, and grab your !PIZZA!
Tomorrow (actually later today since i am way up past my american bedtime) i will post some of my 20+ days worth of content. Thank your for the motivation, inspiration, and encouragement. Hopefully, when AI takes over... it will consider all of us experimental artists as some of it's first and best friends and perhaps it will spare us from the impending SkyNet doom! hahahhaha
Thanks for the....
Haha yeah, better we start to praise SkyNet right now, before is too late!
Cool, I'm eager to see what you deliver "tomorrow" so "my evening" hehe
I'm glad if I'm somehow inspiring hehe thank you!
Get some more !PIZZA!
By the way, I read a comment of you somewhere, and cannot find it anymore lol... so I'll reply here.
It is about, trying to make videos and getting errors. You have mentioned that got a K-80 and were trying to make a 720x480 video, if I'm not wrong. That's impossible, k-80 cannot bear more than 480x480 pixels... so to keep the same aspect ratio you wanted for the video 3:2, you should render it at 480x320 (maybe K80 could bear 540x360px, but I'm not sure), and then upscale it... the good notice, is that as the videos are images in motion, any blur that the upscaling could cause, gonna be far less noticeable than in a static image.
Hey, thanks for the info. I was setting the videos to 640x480 and that may still be too big. Some of the parameters i am not familiar with so it's either operator error from ignorance or the code at times. I am sure it's probably almost always my misunderstandings or bad inputs.
I also ended up taking the day off from internetting yesterday... i was getting digital burnout... so i will work up on the post today a bit. Thanks again my friend. =)
Yeah, sometimes a break is a must, I got a fried one week ago, you wisely warned me about, thanks! Now I enjoy more, and stress less (:
Are you using the video with features notebook? If your problem was that the resolution of the videos was too big, then it dropped a CUDA error. If not, anything else might be related to the prompt. Had to figure that respect the format of the given example is a must, the regular way that we use to prompt in the other notebooks, breaks the code of this particular one.
Take care friend (:
@jotakrevs i just tried to run the video notebook and upon installing the Libraries i got this code...
Downloading CLIP...
Downloading Python AI libraries...
Installing libraries for handling metadata...
Installing Python video creation libraries...
NameError Traceback (most recent call last)
<ipython-input-4-3acbd96b0750> in ()
20 print("Installing Python video creation libraries...")
21 get_ipython().system('pip install imageio-ffmpeg &> /dev/null')
---> 22 path = f'{working_dir}/steps'
23 get_ipython().system('mkdir --parents {path}')
24 print("Installation finished.")
NameError: name 'working_dir' is not defined
I see in step 22 is where the error occurs. Do you have any idea what that error means or how i might go about correcting it?
These are SO amzing and the colours are superb. I had started a series of people merged with tech the change of human as tech becomes more a part of our lives cannot but help inspire art. Really wonderfully moody and true emotion off of these pieces thank you for sharing them @jotakrevs
Thanks a lot for your kind comment, and sorry for the late reply. Didn't get a notification, or maybe it got lost between the trail votes.
I'm glad you liked them, every time a real artist likes my stuff made with AIs I feel very happy!
Regards and take this !PIZZA, is for you!
Fantastic portraits!
Thank you, I'm glad you liked them!
!PIZZA
Congratulations @jotakrevs! You have completed the following achievement on the Hive blockchain and have been rewarded with new badge(s) :
Your next target is to reach 50 posts.
You can view your badges on your board and compare yourself to others in the Ranking
If you no longer want to receive notifications, reply to this comment with the word
STOPTo support your work, I also upvoted your post!
Check out the last post from @hivebuzz:
¡Felicitaciones!
1. Invierte en el PROYECTO ENTROPÍA y recibe ganancias semanalmente. Entra aquí para más información.
3. Suscríbete a nuestra COMUNIDAD, apoya al trail de @Entropia y así podrás ganar recompensas de curación de forma automática. Entra aquí para más información sobre nuestro trail.
4. Creación de cuentas nuevas de Hive aquí.
5. Visita nuestro canal de Youtube.
Atentamente
El equipo de curación del PROYECTO ENTROPÍA
PIZZA Holders sent $PIZZA tips in this post's comments:
jotakrevs tipped aemile-kh (x1)
jotakrevs tipped donnadavisart (x1)
@jotakrevs(1/10) tipped @eve66 (x1)
jotakrevs tipped castleberry (x2)
jotakrevs tipped dbddv01 (x1)
jotakrevs tipped lavista (x1)
Learn more at https://hive.pizza.
These just look so very good! The colours are amazing too.
Thank you for all the effort you are putting in to get into the mind of the AI and sharing your results with us :)
I'm glad you liked them, at least fry my brain is giving some interesting results hehe
Regards and grab your !PIZZA!
Thanks for the Pizza :). So, again, does AI generated art, when used as initial image, give more consistent results?
It does, but the initial image needs some work on it, depending of what you want. Portraits are tricky, so had to made all that I told in my post. Landscapes seems less cruel, at least in the few tests I made.
Yes, I too feel when initial images are used from artbreeder for landscapes, the results are gratifying. I haven’t really tried a lot of portraits, the times I tried, like you said, different parts would get messed up. I used to have a lot of trouble with the eyes especially. I will apply your methods and try again!
Yes, the eyes seemed the worst to me, but I think because they are eyes it becomes more noticeable... my methods worked to fix them enough well, and then is when I noticed that the noses and ears gets totally destroyed lol And as you can see in mi images, ears and noses still a mess, I don't know how to fix them.
Best advice I can give to you, is try to use stylish images, alike drawings, in grey scale, the narrower dynamic range, the better. And have fun!
Yes, fix one problem and you notice others, haha! thanks for all the advice :)
Esto es absolutamente asombroso, los colores y los detalles.
Muchas gracias! No te olvides tu !PIZZA!