What Is Your Ai Art Workflow?

Members

Keep this to answering the question only. No 'is AI good and bad' conversation welcome.

Thank you

____________

Abbie Shores
Manager FAA/ Pixels

My workflow: Pour cup of coffee, sit down in front of computer, pull up AI site, type in a few words, press button, post result.

Depends where you do it, sites have credits and experimenting with words takes time.

I use easy diffusion on my machine got a video card I could afford and test models all the time, I don't in fact make things as I simply test models and shave off the images I like. Got to test lora's too.

Once you find a model you like - civitai.com is where you can get them, decide if you can only do a v1.5 model which only lets you do small things. Or an SDXL file so you can make larger things but the model is huge.

I have a test sheet of a 140 prompts testing the different interests I have, choose the best ones and sort them out, labeling what their strongest things are in the file.

From there you have to test what the best words are, as while you can just hit enter - without any words at all and it will make - -- something--- you still want some control. So you have to figure out the exact wording, and this is where it get hard as you have to take apart the scene and list everything..

A transparent jar with oil and herbs in it. Works, a clear jar may not work. Prompting is hard, though they have a new version of stable diffusion that makes prompting easier. But its still a pain... Anyway, I made the image, look up close, do they look like someone taped a pencil to a cat laced with speed? All squiggles and terrible, that's a fail. Does it look clean and nice? Good. Does it look artististic? Good.

Then I edit like usual, cloning stupid parts out, odd things, I have to smear a lot back into shape, add digital parts because it can't make a clock or a gauge, remove words as it tries to show me its sort of learning the language by adding dream words. Remove watermarks, because its taken from other artists and it adds its own word that looks like a watermark, but still needs removing. Oh and I have to enlarge it before that, and basically that's it, its still a days work and a lot of sorting through junk. And I still want it to look like something I would have made.

If you do it at home you don't have to worry about censors and restrictions and credits. But you do need a good Nvidia card with many cuda cores and as much memory as you can afford. You won't be able to do some models if you don't have at least 12gb, you can cut it with 8gb sometimes. But it will be out of memory pretty fast.


----Mike Savad

Thank you, Mike, this is very helpful!

I will have to check my computer specs when I get home, it was originally bought for gaming so it might actually be able to do the job.

I am definitely interested in doing something like you, where I take the AI image and then make edits and corrections. I have some digital art experience, so I really love the idea of using AI as a starting point and then still putting some of my own work on top of it.

Do you happen to know of a website/service that is high quality and/or that can run different models that you would recommend? In case I am unable to run the models on my own machine.

All I know is the radeon I bought did not work. I think they now have ones that can work with it, but oh well.. It does have 16gigs of memory which is why I got it. Also make sure your pc has enough power to run these things, a stock power supply often isn't enough to run these things.

I do know civitai does allow making of things if you have some kind of points or something. I never used any of these online. The models I choose may not reflect the work you want to have. Each one is specialized to the thing you want. Monsters, creatures, architecture, cars, painting styles, artist rip off styles (mostly in loras), etc. I've rejected hundreds of models, narrowed it down to a dozen, and even then i'm not sure which one i'd use daily if I ever actually use this the way it was designed. Only a few stand out. The rest are usually merged models with more merges. Or worse, trained models using AI images as a model which messes it up more.


----Mike Savad

Also note that the size of the image will alter the contents a lot, small, large, long etc, the images will always look different. Also save all prompts/models/seeds step etc so you can repeat it. And if you use it at home look into Xformers, each diffusion thing has its own and it will speed things up at the cost of the image not looking exactly the same on the same seed.


----Mike Savad

"I really love the idea of using AI as a starting point and then still putting some of my own work on top of it".

That is exactly my preference also. While it's nice to get images occasionally that require no additional work, I feel that creatively, it's a lot more satisfying to add your own touches that (hopefully) turn the start image into something more interesting, and perhaps a bit different to what other people are producing.

I use Stable Diffusion / Automatic 1111 on my own personal computer which is great because it's all free and you can produce as many images as you like. As Mike mentioned, you do need a reasonable spec PC, but if you can do it, it's worth it.

I discovered a program called Studio Artist about a year ago, and that's what I use to add my own creative touches. Then to finish, I often use an image editor for things like clean up work, brightness / contrast, and colour adjustments. It's not a quick work flow, but because it's part of the creative process (and I am still climbing the learning curve on these programs), it's all enjoyable.

Good luck with your future AI endeavours.

To obtain something that I truly like requires significantly more effort than simply typing in a few words and pressing some buttons. I would estimate that on any given day, it takes 50-100 attempts to find one I like.

From there, it is a matter of tweaking with PhotoShop, Tozaz, and or one of the other editors.

The thing I would caution is I have nearly 300 images uploaded here, and they are not selling.

I have displayed some of my AI artwork in local venues, but they didn't receive much attention. Only a couple of my abstract wine pieces were sold in a wine restaurant that claims to be the largest wine seller in the area. Surprisingly, it has been one of my best-selling venues for more than 8 years of selling my traditional photographs.

I have tried nearly all of the generators, but my two favorites are Night Cafe and Microsoft Image Creator. For Night Cafe, you have to buy or earn credits, while Microsoft IC is free for now.

There are lots of channels on YouTube about AI image creation, and many of them are focussed on technical aspects, but if you are new to AI and interested in ways AI apps can be used to produce interesting, creative, and artistic images, then I would suggest looking here:

https://www.youtube.com/@MakingPhoto/featured

The person that runs the channel is a professional photographer, and also a woman!! I only mention that because her videos are great for how to create artistic / creative / interesting images, while most of the men's YT channels are great for geeky things, but not so good (IMO) for showing how to create artistic / creative / interesting images! Hope that's not too controversial for the discussion forum :-)

Not once have i managed to get an image straight out of ai that looks right. I just can't seem to do it. So instead i take my photos and then run them through ai. I then have to put them in either Photoscape or Corel, and then play with it more. I wish i could get good images just from text, but I'm strongly suspicious of people who say they can now hahaha. Besides, I'd rather use my camera and then play... Seems less cheaty somehow, although i know people who CAN do that are like authors of the image. I'm actually quite jealous

"Not once have i managed to get an image straight out of ai that looks right."

I have managed that occasionally, but not very often. Personally, I think some people's expectations are far too high about what AI can actually do. Compared to about a year or so ago, it has made considerable progress, and no doubt will continue to do so, but personally, I think the quality of current AI images is a bit like the early days of digital photography.....good, but not good enough (at this moment in time).

Models make a huge difference in how well they work. I did one where I wanted a steampunk something, and it gave me a train. It was terrible. But certain models do really well. Like if your doing this at home:

socapunk_tech an art photo look
Ultima (needs a vae) a fairy tale look
TheAllysMix good for illustration
Painters Checkpoint - is good for an oil painting look
RandomMaxx is good for photos
Aderek is for an artful photo, its just more arty looking
NewRealityXLAllInOne - worked best for Anthro stuff


and I have some second tier stuff if I can't get it to do what I want. SocapunkTECH (there are a few of them), and Painters Checkpoint - follows prompts very well. More the painters one though. The painter one reminds me of DA painter, has that same look.

The word choices really do matter, a charming cottage in the country of france will turn out differently if you change the country to finland and it changes the architecture. A raw photo will give you a photographic look, where as not saying it may randomly make it a cartoon. Describing it just what it looks like is hard.

Describing things is like this video:



what seems obvious in your head isn't so much so you have to explain what you want and hope the model has the parts you want. The setting, location, lighting you want, angle you want, objects you want, expression etc needs to be outlined, or you'll get a random thing... Well you will still get a random thing. A birds eye point of view, a wide angle lens, the exact camera, a viewpoint from a mouse, a view point from sitting on the couch, will change the angle. Unless you use a lora to force it to look a certain way, naming the style of architecture is needed. Can't just say its from the 1930's, its art deco.

Still telling it not to put a car into a victorian theme is hard. And over explaining messes it up. It likes to be creative so if you add too many words it can confuse it. But you can also add weights to things a glass on a rough table with a (pitcher of lemonade:1.0) to add weights to it, can force something. Each diffuser uses its own thing like ((())) brackets to add or remove things. Which sometimes works. The weights are better.

Where people come up with words I don't know. But my fav word, and there is a lora.. Is GREEBLES.

That fun word is used in the movie industry to add junk the outside of things to add texture, dimension and scale. Like the millennium falcon has all those lumps attached, those are greebles. And just adding that word will often add a lot of extra stuff to a scene, more pipes, more bottles, more clutter that makes it seem more real.


----Mike Savad

That is interesting, Abbie. I have seen some of your creations from your photos, and they are fantastic—at least, that is what I think I am looking at.

Every time I put one of my images into AI, I seem to get something weird out of it.

I guess it's the old saying, garbage in, garbage out! lol

Because I haven't sold many AI images at all, I am sort of hung up on the idea of getting some of them sold just for the challenge of it. lol

Greebles? Interesting...

I've not sold one, Floyd. But then i rarely sell anything as I'm too busy marketing other people and sites lol

I use Midjourney. I think it has the best artistic output, albeit that comes at a premium.

I upscale to print quality with Topaz PhotoAI and edit in Adobe Photoshop.

I think of AI art creation as an explorative process. It's not about perfectly controlling what you get, it's about providing the right inputs and and the right amount of chaos, generating lots and lots of images, and curating out and improving the best.

I've generated over 60,000 images, I have less than 300 on my page.

I've sold 2. But lately I haven't sold anything. Its both encouraging and discouraging that it sells. I have an idea of what should sell but I still have to make a batch, still trying to work out the combo of using a good model with a good lora (There are a lot of really dumb loras), to make it unique. The thing about models is, most have the same exact base model. So across models your description will produce nearly the same results. But there is sometimes that one model that stands above them all. Like the ones I listed. I have others but they are either less creative or don't follow prompts as well but they specialize in one thing or another.

Line AnthroMerge which was is good for animals, but is pretty straight laced boring non creative which makes it good for testing lora's.

Or Pirsus Artstation, which is a v1.5, but it can handle large things, and its super surreal, it does a painting look, but sometimes its like its tripped out on lsd.


----Mike Savad

"Not once have i managed to get an image straight out of ai that looks right."

I think which tools you are using has a lot to do with this.

That being said, so much of curating what pieces I am going to improve is about looking for all the little things that are wrong in an image and deciding if I can and want to fix it.

Midjourney has inpainting now so you can regenerate parts of images and Photoshop's Generative Fill have made this a lot easier though I still find myself hand painting fixes often.

I do a lot of "kitbashing" where I take multiple versions, variations, or upscales of the same image and lay them over each other and use layer masks to blend them into a final image that is free things that bother me.

I haven't sold enough to pay for a month of Midjourney, let alone my yearly subscription for Photoshop, but that's not why I do it anyway.

Mike, perhaps you might be interested in this model?

https://civitai.com/models/216439/artium

Described as "An offshoot from my OMNIUM model, geared more towards an illustrative and painterly style"

Its based on the SDXL 1.0 model.

I am not familiar with the ones you mention above, so will have a look at those.

Mostly I make two or 3 a day. Now sometimes I look them over, and decide I don't like one of them. If I can't fix it then I delete it. When I do landscapes sometimes they need a little touch up. You can use one of several programs for that. Many times I just use paint 3D,but I have paint shop pro now too. It all depends on what needs to be done.
I had one landscape that showed 2 suns. That I just fixed using the paint 3D on my computer. You can do a lot with just a paint program.

Good luck trying out some AI art.

I'll download it, could have sworn I did.. Right now i'm trying

https://civitai.com/models/348244/xl?modelVersionId=389602

which is nearly impossible to search for because its called XL?

It creates a sort of painted/drawn look that looks a cross between digital art, digital painting, anime look. Which i'm not sure if i'll use it as its outside what I normally make, but the houses are nice, and the architecture is wildly different. Does weird things with people portraits (I use 1600x1024) which sometimes does weird things to models.

I check the new models daily as its constantly being updated and added but I can't get them all. I've almost maxed out my monthly comcast amount twice now. I wait till the end just to make sure it doesn't go over 1200gb, 1000gb, didn't think i'd get that high.


----Mike Savad

Oh one test for food is this:

a turkey drumstick on a plate, with an opened hot sweet potato with brown sugar and butter, a pile of corn, rice, and gravy. A gravy boat on the table, place settings.

Sounds simple but I get a mutant bird 95% of the time. Or ask for cornbread, and nearly always you'll get bread with corn wrapped around it.


----Mike Savad

That's a really good test prompt.

https://cdn.midjourney.com/be30ac52-9e11-4fc7-b6c3-3234ebf4585a/0_3.webp

Midjourney did pretty well I think, all the elements are there, not quite an "opened sweet potato." Also it looks like a wine glass of gravy in the background.

This one is hilariously wrong though...

https://cdn.midjourney.com/be30ac52-9e11-4fc7-b6c3-3234ebf4585a/0_0.webp

Oh..a leg of sweet potato! That will suit me well.. Veggie meal lol

It does a good job making it look real. Usually I get something that has 5 drumsticks coming out in all directions. Or a pile of drumsticks

a delicious pile of bbq baby back ribs, smoked and glazed in a nice sauce, on an oval metal plate, extra dipping sauce, a small basket of bread, butter pats, cornbread, corn on the cob, mashed potatoes, a soft drink, product photography, dramatic studio lighting.

This one will give me a mix, almost always i'll have a glass of bbq sauce with a straw in it. I mean the sauce is good, but its not that good. I also try cake which usually works but it doesn't know what a pettifor is. It knows its cake, just not the striped cake. It makes every kind of cake but the one I know of. Unless my cake classification is wrong.


This one is always fun and i did it accidentally

A HIGHLY DETAILED POODLE, OF A ROOSTER, raw photo, professional photo

I forget what I was going for but I usually get a fluffy, 4 legged chicken. But i'll get weird combos either way. Try wording it differently and I could never get it to mix like that. I haven't posted any as its mostly a test to see how the animals look or if they combine weird.

Getting things so the backs are to me, like sitting at a bar, is a lot harder than I want. I wanted animals there and they all almost always pose in a line up like a portrait. I guess they don't have casual models?


Mostly I like painted models as AI tends to look terrible up close.

----Mike Savad

Philip, thank you for the YouTube suggestion, I will definitely check her out!

And thank you to everyone else as well for the replies, I have a lot to digest. :)

I typically use Canva and Ibis paint. My tools are tablet based, so AI is somewhat baked into the system, even when I don't use it.

The process always starts with some kind of an inspiration, and then some diddling around to come up with an image that fits the inspiration. It's not really a straight line process, but a back and forth between my tools as I piece together my image. It's not a one shot and done process but a rather labor intensive process where I combine the components I want and create the image appropriate to the original inspiration.

The most important thing I have found is that your picture must elicit an emotion when you are done. If it feels "lifeless" or "dead", then you have relied too much on the AI and not enough on your inspiration. The technique in many ways is not that different than a traditional medium when used in this form and framework.

It is simply a tool in my tool chest with all the other tools that gets used appropriate to what I am actually working on.

I think the most critical element of the entire process is that by the time you are done, whatever you have created must be distinctly yours and your work, no matter what tools you use or which order you use them in.

+1 f or @Daniel Super's answer: Midjourney. Mike is smart and active member here, so I know he will take this with a grain of salt: if I'd read his "how to" on using AI tools I never would have tried them! Why not? Because I work for myself, so already have to solve problems with the pc, the phone, the router, the printer, software etc. So that LAST thing I want is more tech problems to suss out, I simply want to create.

Midjourney is *hands down* the most creative tool for generation. I mean for coherence, mood, interpretation, color. The others are good or even great at *other things* like reproducing real world physical objects like toasters, cars, spaceships, or for doing strictly political or erotic art (and that isn't going to change soon, especially in an election year with most of the players nervous about bad press for AI in general)

Prompting is a craft, though, like anything else. It is a huge help to know art / photography / cinema concepts and styles. That being said, some of my favorite results have been completely out of the blue, just prompting things like titles of poems, or mashes of things that aren't logical but intuition led me to try, like -- "Arabian horses at daybreak shot on expired film stock by Giotto, lofi luxury vibes" or something similar gave a beautiful result that the literal prompts were completely missing. In short, there is an element of *collaboration* with Midjourney that I find missing in the others. Good luck!

I built my pc for the possible use of AI. But when they said you must have a card of at least XX amount and my card only had 3 that's all I looked for. So I went for the best card I could find up to a certain amount, still went over the budget, got lots of ram. Only to find out that AI can only be used on CUDA, my expensive card while rated well on games, wasn't rated for AI, and wasn't at all compatible. So its very important that if your machine doesn't have the power, or the right card, you won't easily be able to do AI. Though maybe they have the right things now, I haven't checked.

The problem with the online things is they do censor you, control what you are allowed to make and not, its also public from what I tried that one time. There isn't really enough room to experiment with things. By the time you finalized what you want, sorry you ran out of credits. And hopping from one site to another isn't ideal either because the style changes.

I don't know where the models came from from civitai people use or made. But midjourney took a lot of stuff from this site and others, so just out of principal I can't use it. I'm certain the other models came from other people's art but I don't think they did it in such a way as the first generation models did. Not to start that war or anything but I tell it like it is.


----Mike Savad

As has been mentioned, it takes experimentation and several iterations to get it right. (A relative statement)

I start with an idea, write a prompt (at first relatively short and general) and then refine it by adding detail, changing words, characteristics, syntax, grammar etc. until I get zeroed in on something I like. Depending on the result, I'll either save it for later manipulation or trash it. (sometimes hundreds) I've read you get better results being very specific with your prompts. It's a struggle to be specific up front (for me) because I often am not sure what exactly I want to start with, just generalities. Well, I have a range of emotional responses to the results I get from sheer disappointment to absolutely wowed. I have only been using the free BING Image Creator so far and have posted several pieces here on FAA. Those few are the culmination of hundreds of tries and yet, they are not perfect and that's fine.

I never use an Ai image straight out of the box, I always have to modify it somehow. So my Ai routine is becoming clearer as I do more of them but at this stage my "work flow" is not really defined. And, I'm not really sure I would call it "work flow" because that sounds like something it isn't, work. For me, it's literally "playing around."


.

I have never used Midjourney, but my impression from YT channel geeks is that it probably is the best of the bunch currently available (eg, for artistic / accuracy etc). However, the great thing about Stable Diffusion is that its open source code (thank you Stability AI), and has large numbers of people contributing to its development in all sorts of different ways, eg Web UI, models, etc, and while Midjourney might have the upper hand now, how long will that last? Plus Stable Diffusion if loaded and run from your own PC is completely free to use, with no limit on image generations. If I was starting from scratch, I would still choose SD over Midjourney. Just my opinion of course, others may take a different view.

Robert - I agree - it is a labor intensive process. And same for me, a tool in my toolbox.

My work flow starts with coming up with an idea of what I want to create. Sometimes I will use one of my own photographs as the base (I mostly use Night Cafe, though I just recently started working with Corel Paint). Either way, my prompts start with what I would write for alt-text for one of my photographs. That's actually how I got started with AI art - wanted to see what image I could create by using my skills of creating alt-text. I then use my knowledge of art styles and mediums (loved visiting art museums and then during the pandemic a lot of time on their websites). Every model on Night Cafe is different and yields different results, so I tweak my prompts until I get what I like.

As Scot said, prompting is a craft. And yes, it's important to know art or photography concepts and styles. That's where I use my knowledge of photography to achieve the look that I want, using terms like bokeh, depth of field, etc. I also use "negative prompts" in much the same way - to reduce aberrations or issues with what the AI model produces.

My work flow also consists of a lot of reading to learn more about art styles. Rather than invoking an artist's name for the AI model, I try to come up with the words that would accurately describe that artist's style or art movement, like impressionist, surrealism, gauche, etc.

I'd like to "throw down the gauntlet" -- artistically speaking -- and do a *visual challenge* as to what the different softwares are capable of. I don't know if there is any easy way. Forums don't allow image posts (methinks) and there doesn't seem to be a collective gallery where we could all upload to. Or maybe I've missed it??

Most of my best work I've kept for myself, I've seen (and been discouraged by) what people spend money on, so I haven't bothered to upload the more interesting results. Collages, double exposures, film stills from non-existent movies, even funny stuff that I would post to the "Make me laugh" thread if I could....

Hmmmmn....

PS

https://abbie-shores.com/pixels

Most members use my special host

Well, they sure won't sell if you never show them.

I make a big mistake when I try to second-guess the marketplace and predict what will and will not sell. I upload everything and trust that people can make up their own minds about what they want to buy or not.

@the wow - its not just the software but the model, each one is very different. There is no easy way to tell. Some are very sharp, others are very horrible. Its why I rate them on a score sheet under 24 factors. Many don't pass.

If it looks good post it, as long as you can give it a good title, description etc. Some are so "out there" that its hard to put a label on.


----Mike Savad