Embedded Files
This image was created woth the Goddess prompt in a more recent engine. I will explain a few things I learned making human figures in this post. With prompts.
The human figure on early generative algorithms
As you might recall (any of you still here) the main algorithms used for generative art were distant relatives of Deep Dream. Things that used Genarative Adversarial Networks (GANs) and, a bit latter on Contrastive Language-Image Pre-training). There were many different ones and it was kind of a Cambrian Explosion with no clear winner... up until the Stable Diffusion and its kind.
I tested many different generators and the best results in this era I found to be given by the two engines at NightCafe. The engines were called Artistic (a GAN+CLIP algorithm) and Coherent (CLIP+Diffusion). They were very nice to work, one usually alternated images between both and produced very nice things. Most of my first Sketchbook was generated this way (plus GOF digital editing, off course). Still today, If I want to generate a pretty landscape I use then to generate the starting point for Stable Diffusion.
These older algorithms rock at landscapes! They also create a very pretty, artsy texture on other things.
Enough telling, lets show. ...
One of my favorite images of the first SketchBook was elven departure.
In this image I was trying to use Stable Diffusion to make an image of elves using vehicles. I wanted a bucolic feeling too, with wind blown grass and stuff... So the prompt went;
"machine of the elves moved by the wind through the grass, an elven walks along in the first plane...
It was... lets say, not spectacular.
In fact evet to get this result I had to use image to image because the landscape was dull when I tried to use Stable Difusion from scratch. When I first tried to make the wind machine to add the elven latter (someting that is usefull in Stable Diffusion but requires outside editing) the scene was ... not good and the machine was rather unnimaginative.
So I used a seed image created with GAN (on the side) and the same prompt.
As you can see the main elements are all there, as well as the colors are much better, but there is no decent coherent image yet, more like a bunch of blobs in the right places for the final image.
Still, using Stable Difusion on this image got me the dull one... so, what to do?
At this point I started using Artistic and Coherent to fine tune the results I wanted. Artistic usually create chaos and color, but is less... well les Coherent. Bellow you can see the evolution from the first Artistic to the final Coherent. Note that I made several cycles to refine this result.
Artistic applyed to the dull image
Following artistic with coherent gets us close to the end result
So, there were still things were Stable Diffusion was not so good (specially if you are trying for a more artistic image). But, on the other side, the same idea of alternating Artistic and Coherent for portraits led to awful results. One example on the side.
Off course, they can be better using a starting photo or drawing, but that is not the point. The point is that Stable Diffusion gave us a tool for portraits.
Now we had Stable Diffusion and, from the very begining, it was very good at the human figure. Not so in everything else (off course it changed and now it is pretty much the only game in town).
Now we had Stable Diffusion and, from the very begining, it was very good at the human figure. Not so in everything else (off course it changed and now it is pretty much the only game in town, for now).
But keeping on with my narrative: I started creating series of portraits to test models, poses and prompts. In fact I continued this work afterward and still do every time the tech changes (probably will still do it once ControlNet is more widelly available). Many of the results turn out NSFW, which, for anybody that does learn to paint is quite natural but can be daunting in the generative art landscape due to censorship embedded in the tools*.
The first prompt I made with good results is what I called the Vogue Frazetta prompt. It gave good results on NightCafe and anywhere else I tested. I will share this one with you:
beautiful long haired female model, a vogue photoshoot, full body image, anatomically detailed, golden chains clothing, art noveau clothing, painted by Frank Frazetta, hyperdetailed, beautiful, fashion, trending on artstation, oil painting, digital art
Image from scratch, very simple, no negative prompt nothing fancy (prompt strenght at 50% - NC, or 7/8 - PlaygroundAI etc). It does work. It also generates lots of NSFW things that sometimes are censored sometimes are not.
You can change two parts: the clothing and the "long haired" part to generate different kinds of images. Without the negative promot there will be many cropped images, so it is advised to use, even the one I sugested bellow will be good.
As I said these prompts were not easy to come by so I kept looking for good ways to improve the portraits and make less mistakes. I will share some of these with you also. But going on chronologically the second major prompt came from a series of ideas I got from other prompters.
It is called Goddesses and it still is quite good. I must remind you that those were prompts using Stable Difussion withour ANY negative prompting. Using then in more recent versions of SD and with adequate negative prompting will give you clearly better results.
I will also share this prompt with you **:
very beautiful sensual flexing fire goddess *supressed** with jewels and lots of tattoos. white skin. Long nails Long flowing red hair. intricate and hyperdetailed matte painting by Huang Guangjian, Anna Dittmann and Dan Witz, fantasy art, loving, Unreal Engine 5 polished
The suppressed part is an expression that is now mostly resticted, but it works without it. You can add some detail you want there and also and you can change the description. This one require a stronger prompt weigth, something like 80%-90% on NC. Testing it with less then 14 in PlaygroundAI will not give you etnic results so you may have to go very far (as far as 20).
The one from bellow, for instance use different hair, skin and details and is my favorite. Like Vogue Frazetta, it does recquire post processing.
Now, this third prompt is also something I picked up and improved over time. I like specially how it give very artistic textures. I call it Nerdy from the first image I made with it. It is a very versatile prompt and there are traces of it in many other works and prompts I made afterwards. Here it is:
Most beautiful nerdy woman :: ***how she is - expression*** :: ***how she is whole body clothing *** :: full body/full head ::Sultry, attractive, provocative, perfect proportions :: portrait by Daniel F Gerhartz :: fantasy art :: sharp clear, photo realistic, hyperrealistic, hyper detailed
There are three parts you can configure in this one: the overall expression (anguished, happy, distressed, etc...), the overall way she is dressed (torn clothes, formal clothes, dirty clothes, ...) and the framing of the portrait (full body, full head, head and shoulders, wais to head,...)
Now a few other things I picked up and will share with you that help you making a better prompt for protraits (and a better prompt in general).
Negative prompt is fundamental. A good one will improve your results dramatically. NightCafe uses now a default one (you can change it and disable it). Many other tools also have a embedded one but they do not show it. You can still add more things with negative weights, will just have to find out the sintax for weighting in the prompt for that particular platform.
How tools like Midjourney, Lexica, DALLE do pre-prompting and choose part of your prompt for you is probably part of their secret sauce, so I have no clue. But they are all variations of Stable Diffusion and the more open tools all do pre-negative prompting (or allow you to) so it is a sure bet they do something along these lines.
If you are out of ideas for a negative prompt try this one (a 0.5 weight for a positive 1.2 is my favorite variation): " ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, blurry, bad anatomy, blurred, watermark, grainy, signature, cut off, draft"
It is important to figure out how your tool treats long prompts cause it change a lot of the effects. Some tools restrict the length of the prompt, but allow many prompt lines with weigths. Some others do not restrict but just ignore the final parts. Some can understand very, very long promots quite well.
There is a tendency to keep hoarding words ina prompt that are just unused. Check it from time to time and keep you prompts as lean as possible. Always keeps testing and refining your prompts.
The first thing you say should be an overal description of what you want. The object of your generation. Most online generators allow using parentheses ((())) around the focus of your prompt if it is not the first line. Not all online generators parse Stable Diffusion sintax for weighting, but many do. As always, keep making your experiments.
Artists styles have unintended consequences. A few I found out by trying and error (remember these are form MY experience and may be subjective observations, some may be not even be the case anymore):
Luis Royo makes beautiful female designs but calls for foggy images, Seb McKinnon creates beautyful Sandman Style images and goes well with a fairytale look but also a bit foggy.
Luis Royo, Pino Daeni, Greg Rutkovski and Lucas Graciano give beautiful fantasy art but many artifacts (either strokes or blur), John Williams Waterhouse and Tom Bagshaw give a more romantic finish.
Artgerm will try to make anybody you did not properly specify to be a female
Anna Dittmann will focus on face portraits unless clearly specified (many others do it to like Alphonse Mucha)
Frank Frazetta and Moebius do not go well with photorealism (one gives a plastic look, the other jumbles lines)
Alphonse Mucha and Anna Dittmann will give you a frame to your design.
Have fun trying different styles and combining these prompts with others. For instance, since these were not focused on photorealism you can add parallel prompts on photorealism and find you combination.
Share your findings, there are always foruns and the community is quite nice in most of them.
That is it. I will put bellow some of the results of these prompts as thumbnails. I may post the full versions on a more controlled platform like Artstation or DeviantArt. In some more family friendly cases maybe even Instagram.
Some links for the most curious among you:
I - Vogue Frazetta
II - Goddess
III - Nerdy
IV- Nerdy++
IS IT THE END?
Off course it is not. I kept working on this and there are a lot of thins to say about some other desirable skills while creating a portrait:
Photorealism, Stable Diffusion is not MJ5 but there are lots of things you can do to achieve better photorealism.
Posing. When you creat a story you may want to construct a scene in the right way. While there is a full solution to it (ControlNet) this is not available online (at least now in April 2023). You can download SD and install it for free but if falls on the "too much commitment for a beginner" thing. But you can do it, with some restrictions without ControlNet, I will talk about it in the future.
Character consistency is another problem if you are interested in storytelling so it is another thing I should talk about.
I will keep talking about what you can do without installing tools (you can do a lot more if you run SD locally) , but I assume this is too great a commitment for those starting.
SOME FINAL CAVEATS
*There is a whole lot about what these tools think is NSFW and censor and how their criteria make little sense for most of the world outside the US, but this is a subject for an entire post.
** I had to make an adjustment, the original used one word that is now prohibited. This may happen with some prompts from time to time and from platform to platform. For isntance, words like buxom and bare cam be prohibited in some systems. Others, like nude can be used in the NEGATIVE prompt in some systems (which is useful, believe me) but not in others,
*** I really find annoying how many people only use Generative Art to create lots and lots of hypersexualized Waifus. Nudes and eroticism are always a part of art but I am talking about unimaginative-waifu-after-waif with increasing degrees of kink. If it is your stuff, fine. It is not mine, so I will not talk about it, It is, however, the least common denominator and boringly easy to achieve. I am concearned mainly in creating scenes for stories so photorealism, character consistency and posing are my current concerns. These are topics I will cover fin the future..
I - Vogue Fraet
Page updated
Google Sites
Report abuse