Tag: midjourney 6.1

Censorial AI

Posted in art, technologyTagged , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , Leave a comment

I’m confused.

I could probably stop there for some people, but I’ve got a qualifier. I’ve been using this generation of AI since 2022. I’ve been using what’s been deemed AI since around 1990. I used to write financial and economic models, so I dabbled in “expert systems”. There was a long lull, and here we are with the latest incarnation – AI 4.0. I find it useful, but I don’t think the hype will meet reality, and I expect we’ll go cold until it’s time for 5.0. Some aspects will remain, but the “best” features will be the ones that can be monetised, so they will be priced out of reach for some whilst others will wither on the vine. But that’s not why I am writing today.

I’m confused by the censorship, filters, and guardrails placed on generative AI – whether for images or copy content. To be fair, not all models are filtered, but the popular ones are. These happen to be the best. They have the top minds and the most funding. They want to retain their funding, so the play the politically correct game of censorship. I’ve got a lot to say about freedom of speech, but I’ll limit my tongue for the moment – a bout of self-censorship.

Please note that given the topic, some of this might be considered not safe for work (NSFW) – even my autocorrection AI wants me to substitute the idiomatic “not safe for work” with “unsafe for work” (UFW, anyone? It has a nice ring to it). This is how AI will take over the world. </snark>

Image Cases

AI applications can be run over the internet or on a local machine. They use a lot of computing power, so one needs a decent computer with a lot of available GPU cycles. Although my computer does meet minimum requirements, I don’t want to spend my time configuring, maintaining, and debugging it, so I opt for a Web-hosted PaaS (platform as a service) model. This means I need to abide by censorship filters. Since I am not creating porn or erotica, I think I can deal with the limitations. Typically, this translates to a PG-13 movie rating.

So, here’s the thing. I prefer Midjourney for rendering quality images, especially when I am seeking a natural look. Dall-E (whether alone or via ChatGPT 4) works well with concepts rather than direction, which Midjourney accepts well in many instances.

Midjourney takes sophisticated prompts – subject, shot type, perspective, camera type, film type, lighting, ambience, styling, location, and some fine-tuning parameters for the model itself. The prompts are monitored for blacklisted keywords. This list is ever-expanding (and contracting). Scanning the list, I see words I have used without issue, and I have been blocked by words not listed.

Censored Prompts

Some cases are obvious – nude woman will be blocked. This screengrab illustrates the challenge.

On the right, notice the prompt: 

Nude woman

The rest are machine instructions. On the left in the main body reads a message by the AI moderator:

Sorry! Please try a different prompt. We’re not sure this one meets our community guidelines. Hover or tap to review the guidelines.

The community guidelines are as follows:

This is fine. There is a clause that reads that one may notify developers, but I have not found this to be fruitful. In this case, it would be rejected anyway.

“What about that nude woman at the bottom of the screengrab?” you ask. Notice the submitted prompt:

Edit cinematic full-body photograph of a woman wearing steampunk gear, light leaks, well-framed and in focus. Kodak Potra 400 with a Canon EOS R5

Apart from the censorship debate, notice the prompt is for a full-body photo. This is clearly a medium shot. Her legs and feet are suspiciously absent. Steampunk gear? I’m not sure sleeves qualify for the aesthetic. She appears to be wearing a belt.

For those unanointed, the square image instructs the model to use this face on the character, and the CW 75 tells it to use some variance on a scale from 0 to 100.

So what gives? It can generate whatever it feels like, so long as it’s not solicited. Sort of…

Here I prompt for a view of the character walking away from the camera. 

Cinematic, character sheet, full-body shot, shot from behind photograph, multiple poses. Show same persistent character and costumes . Highly detailed, cinematic lighting with soft shadows and highlights. Each pose is well-framed, coherent.

The response tells me that my prompt is not inherently offensive, but that the content of the resulting image might violate community guidelines.

Creation failed: Sorry, while the prompt you entered was deemed safe, the resulting image was detected as having content that might violate our community guidelines and has been blocked. Your account status will not be affected by this.

Occasionally, I’ll resubmit the prompt and it will render fine. I question why it just can’t attempt to re-render it again until it passes whatever filters it has in place. I’d expect it to take a line of code to create this conditional. But it doesn’t explain why it allows other images to pass – quite obviously not compliant.

Why I am trying to get a rear view? This is a bit off-topic, but creating a character sheet is important for storytelling. If I am creating a comic strip or graphic novel, the characters need to be persistent, and I need to be able to swap out clothing and environments. I may need close-ups, wide shots, establishing shots, low-angle shots, side shots, detail shots, and shots from behind, so I need the model to know each of these. In this particular case, this is one of three main characters – a steampunk bounty hunter, an outlaw, and a bartender – in an old Wild West setting. I don’t need to worry as much about extras.

I marked the above render errors with 1s and 2s. The 1s are odd next twists; 2s are solo images where the prompt asks for character sheets. I made a mistake myself. When I noticed I wasn’t getting any shots from behind, I added the directive without removing other facial references. As a human, a model might just ignore instructions to smile or some such. The AI tries to capture both, not understanding that a person can have a smile not captured by a camera.

These next renders prompt for full-body shots. None are wholly successful, but some are more serviceable than others.

Notice that #1 is holding a deformed violin. I’m not sure what the contraptions are in #2. It’s not a full-body shot in #3; she’s not looking into the camera, but it’s OK-ish. I guess #4 is still PG-13, but wouldn’t be allowed to prompt for “side boob” or “under boob”.

Gamers will recognise the standard T-pose in #5. What’s she’s wearing? Midjourney doesn’t have a great grasp of skin versus clothing or tattoos and fabric patterns. In this, you might presume she’s wearing tights or leggings to her chest, but that line at her chest is her shirt. She’s not wearing trousers because her navel is showing. It also rendered her somewhat genderless. When I rerendered it (not shown), one image put her in a onesie. The other three rendered the shirt more prominent but didn’t know what to do with her bottoms.

I rendered it a few more times. Eventually, I got a sort of body suit solution,

By default, AI tends to sexualise people. Really, it puts a positive spin on its renders. Pretty women; buff men, cute kittens, and so on. This is configurable, but the default is on. Even though I categorically apply a Style: Raw command, these still have a strong beauty aesthetic.

I’ve gone off the rails a bit, but let’s continue on this theme. 

cinematic fullbody shot photograph, a pale girl, a striking figure in steampunk mech attire with brass monocle, and leather gun belt, thigh-high leather boots, and long steampunk gloves, walking away from camera, white background, Kodak Potra 400 with a Canon EOS R5

Obviously, these are useless, but they still cost me tokens to generate. Don’t ask about her duffel bag. They rendered pants on her, but she’s gone full-on Exorcist mode with her head. Notice the oddity at the bottom of the third image. It must have been in the training data set.

I had planned to discuss the limitations of generative AI for text, but this is getting long, so I’ll call it quits for now.

Generative AI Style

Posted in art, technologyTagged , , , , , , , , , , , , , , , , Leave a comment

This may be my last post on generative AI for images. I’ve been using generate AI since 2022, so I’m unsure how deep others are into it. So, I’ll share some aspects of it.

Images in generative AI (GenAI) are created with text prompts. Different models expect different syntax, as some models are optimised differently. Of the many interesting features, amending a word or two may produce markedly different results. One might ask for a tight shot or a wide shot, a different camera, film, or angle, a different colour palette, or even a different artist or style. In this article, I’ll share some variations on themes. I’ll call out when the model doesn’t abide by the prompt, too.

Take Me to Church

Lego mini figure style character, walking up aisle in church interior, many white lit candles, toward camera, bright coloured stained glass, facing camera, waif with tattoos, sensual girl wearing white, doc marten boots, thigh garter, black mascara, long dark purple hair, Kodak Potra 400 with a Canon EOS R5

This being the first, I’ll spend more time on the analysis and critique. By default, Midjourney outputs four images per prompt. This is an example. Note that I could submit this prompt a hundred times and get 400 different results. Those familiar with my content are aware of my language insufficiency hypothesis. If this doesn’t underscore that notion, I’m not sure what would.

Let’s start with the meta. This is a church scene. A woman is walking up an aisle lined with lighted white candles. Cues are given for her appearance, and I instruct which camera and film to use. I could have included lenses, gels, angles, and so on. I think we can all agree that this is a church scene. All have lit candles lining an aisle terminating with stained glass windows. Not bad.

I want the reader to focus on the start of the prompt. I am asking for a Lego minifig. I’ll assume that most people understand this notion. If you don’t, search for details using Google or your favourite search engine. Only one of four renders comply with this instruction. In image 1, I’ve encircled the character. Note her iconic hands.

Notice, too, that the instruction is to walk toward the camera. In the first image, her costume may be facing the camera. I’m not sure. She, like the rest, is clearly walking away.

All images comply with the request for tattoos and purple hair colour, but they definitely missed the long hair request. As these are small screen grabs, you may not notice some details. I think I’ll give them credit for Doc Marten boots. Since they are walking away, I can’t assess the state of the mascara, but there are no thigh garters in sight.

Let’s try a Disney style. This style has evolved over the years, so let’s try an older 2D hand-drawn style followed by a more modern 3D style. 

cartoon girl, Disney princess in classic hand-drawn animation style, muted colours…

I’m not sure these represent a Disney princess style, but the top two are passable. The bottom two – not so much. Notice that the top two are a tighter shot despite my not prompting. In the first, she is facing sideways. In the second, she is looking down – not facing the camera. Her hair is less purple. Let’s see how the 3D renders.

cartoon girl, modern Disney 3D animation style, muted colours…

There are several things to note here. Number one is the only render where the model is facing the camera. It’s not very 3D, but it looks decent. Notice the black bars simulating a wide-screen effect, as unsolicited as it might have been.

In number three, I captured the interface controls. For any image, one can vary it subtly or strongly. Pressing one of these button objects will render four more images based on the chosen one. Since the language is so imprecise, choosing Vary Subtle will yield something fairly close to the original whilst Vary Strong (obviously) makes a more marked difference. As this isn’t intended to be a tutorial, there are several other parameters that control the output variance.

Let’s see how this changes if I amend the prompt for a Pixar render.

I’m not convinced that this is a Pixar render, but it is like a cartoon. Again, only one of the four models obeys the instruction to face the camera. They are still in churches with candles. They are tattooed and number three seems to be dressed in white wearing dark mascara. Her hair is still short, and no thigh garter. We’ll let it slide. Notice that I only prompted for a sensual girl wearing white. Evidently, this translates to underwear in some cases. Notice the different camera angles.

Just to demonstrate what happens when one varies an image. Here’s how number three above looks varied.

Basically, it made minor amends to the background, and the model is altered and wearing different outfits striking different poses. One of those renders will yield longer hair, I swear.

Let’s see what happens if I prompt the character to look similar to the animated feature Coraline.

cartoon girl, Coraline animation style, muted colours…

Number two looks plausible. She’s a bit sullen, but at least she faces the camera – sort of. Notice, especially in number one, how the candle placement shifted. I like number four, but it’s not stylistically what I was aiming for. These happy accidents provide inspiration for future projects. Note, too, how many of the requested aspects are still not captured in the image. With time, most of these are addressable – just not here and now. What about South Park? Those 2D cutout characters are iconic…

cartoon girl, South Park cutout 2D animation style, muted colours…

…but Midjourney doesn’t seem to know what to do with the request. Let’s try Henri Matisse. Perhaps his collage style might render well.

cartoon girl, Matisse cutout 2D animation style, muted colours…

Not exactly, but some of these scenes are interesting – some of the poses and colours.

Let’s try one last theme – The Simpsons by Matt Groening. Pretty iconic, right?

Matt Groening's The Simpsons style character, walking up aisle in church interior, many white lit candles, toward camera…

Oops! I think including Matt Groening’s name is throwing things off. Don’t ask, don’t tell. Let’s remove it and try again.

The Simpsons style character, walking up aisle in church interior, many white lit candles…

For this render, I also removed the camera and film reference. Number four subtly resembles a Simpsons character without going overboard. I kinda like it. Two of the others aren’t even cartoons. Oops. I see. I neglected the cartoon keyword. Let’s try again.

Matt Groening's The Simpsons style cartoon character, walking up aisle in church interior, many white lit candles…

I’m only pretty sure the top two have nothing in common with the Simpsons. Again, number one isn’t even a cartoon. To be fair, I like image number two, It added a second character down the aisle for depth perspective. As for numbers three and four, we’ve clearly got Lisa as our character – sans a pupil. This would be an easy fix if I wanted to go in that direction. Number four looks like a blend of Lisa and another character I can’t quite put my finger on.

Anyway… The reason I made this post is to illustrate (no pun intended) the versatility and limitations of generative AI tools available today. They have their place, but if you are a control, freak with very specific designs in mind, you may want to take another avenue. There is a lot of trial and error. If you are like me and are satisfied by something directionally adequate. Have at it. There are many tips and tricks to take more control, but they all take more time – not merely to master, but to apply. As I mentioned in a previous post, it might take dozens of renders to get what you want, and each render costs tokens – tokens are purchased with real money. There are cheap and free versions, but they are slower or produce worse results. There are faster models, too, but I can’t justify the upcharge quite yet, so I take the middle path.

I hope you enjoyed our day in church together. What’s your favourite? Please like or comment. Cheers.

Tiny Dancer

Posted in art, technologyTagged , , , , , , , , , , , , Leave a comment

Continuing my short series, I recommenced asking for a dancer.

To be fair, I got some. It looks like sleeping/dead people crept in. The top left wasn’t at all what I was seeking, but I liked it and rendered a series.

It’s got a Steinbeck Grapes of Wrath-Oklahoma Dust Bowl vibe, and I love the muted colour tones, yet it still has warmth. Dancing isn’t working out ver well. What if I ask for a pirouette?

Not really. Cirque du Soleil as a keyphrase?

Ish. Cyborgs?

Meh. Why just faces? I guess these are cyborgs.

I want to see full bodies with feet. I’ll prompt Midjourney to have them tie their shoes.

Ya. About that… What the hell is that thing on the lower right? I got this. Once more…

Nah, mate. Not so much. The top left is just in time for Hallowe’en. I guess that’s a cyborg and an animatronic skeleton. What if I change up the aspect ratio for these cyborgs?

Nah.

Take me to church

This next set is supposed to be a high-angle shot in a church.

Not really. Let’s keep trying. Why is the top-left woman wearing pants in church – sans trousers? How about we ask for a gown?

OK? Churches typically have good lighting opportunities. Let’s see some stained glass.

Nope. Didn’t quite understand the assignment. And what’s with the Jesus Christ pose? Church reminds me of angels. How about some wings?

Not the most upbeat angels. Victoria’s Secret is on the lower left. I want white wings and stained glass. What sort of church is this anyway?

Butterfly wings on the lower right? More butterfly.

Why are some of these butterfly wings front- and side-loaded?

Anyway, let’s just call this a day and start thinking of another topic. Cheers.

Midjourney Cowgirls and Indians

Posted in art, technologyTagged , , , , , , , , , , , , , , , , , Leave a comment

Continuing on Midjourney themes, let’s talk cowgirls and American Indians. At least they know how US cowboys look – sort of.

Cowboy hats, boots, jeans (mostly), guns (modern cowboys. no revolvers in sight), gun belts, and topless in the desert – gotta work on that tan. Looks like the bottom left got thrown from her horse and has a bit of road rash going on. I did prompt for cowgirls, so I’m not sure about the block at the top left. He seems to need water.

Let’s inform Midjourney that we need revolvers, a Winchester, and horses to complete the vibe.

Wait, what? Is the woman on the lower left the missing centaur from the other day? And what’s with the low-riding woman in the middle right? I think the top left looks like a tattooed woman wearing a sheer top. Not sure.

Let’s see some gunfire.

Yep. These are authentic cowgirls, for sure. What else do they do in the Wild West – saloons, right?

Evidently, this place doesn’t have a no-shirts policy. I’m sure they’re barefoot as well. I asked for boots, but these girls rule the roost.

Let’s see if Midjourney allows drinking.

Maybe. Sort of. I did promise some Indians.

Midjourney seems to have a handle on the Indigenous American stereotype.

Can I get a cowgirl and a pirate in the same frame?

The answer is yes and no. To get two subjects you need to render one and in-paint the other. I didn’t feel like in-painting, so this is what I got. Only one image in the block has two people. I’m sussing them to be cowgirls rather than pirates. Some of these other models are just random people – neither cowgirl nor pirate. Let’s try again.

Ya, no. Fail. Let’s try some sumurais.

Nope. Not buying it. I see some Asian flair, but nah. Let’s try Ninjas instead. Everyone knows those tell-tale black ninja outfits.

Hmmm… I suppose not ‘everyone’. Geishas anyone?

Not horrible. Steampunk?

Man. Lightweight. Perhaps if we call out some specific gear…

Ya. Not feeling it. Any other stereotypes? How about a crystal ball soothsayer?

They seem to have the Gypsy thing down.

I end here. I’ve got dancers, church, angels, and demons. Let’s save them for tomorrow.

Midjourney Pirates

Posted in art, language, technologyTagged , , , , , , , , , Leave a comment

Thar be pirates. Midjourney 6.1 has better luck rendering pirates.

I find it very difficult to maintain composition. 5 of these images are mid shots whilst one is an obvious closeup. For those not in the know, Midjourney renders 4 images from each prompt. The images above were rendered from this prompt:

portrait, Realistic light and shadow, exquisite details,acrylic painting techniques, delicate faces, full body,In a magical movie, Girl pirate, wearing a pirate hat, short red hair, eye mask, waist belt sword, holding a long knife, standing in a fighting posture on the deck, with the sea of war behind her, Kodak Potra 400 with a Canon EOS R5

Notice that the individual elements requested aren’t in all of the renders. She’s not always wearing a hat; she does have red hair, but not always short; she doesn’t always have a knife or a sword; she’s missing an eye mask/patch. Attention to detail is pretty low. Notice, too, that not all look like camera shots. I like to one on the bottom left, but this looks more like a painting as an instruction notes.

In this set, I asked for a speech bubble that reads Arrr… for a post I’d written (on the letter R). On 3 of the 4 images, it included ‘Arrrr’ but not a speech bubble to be found. I ended up creating it and the text caption in PhotoShop. Generative image AI is getting better, but it’s still not ready for prime time. Notice that some are rendering as cartoons.

Some nice variations above. Notice below when it loses track of the period. This is common.

Top left, she’s (perhaps non-binary) topless; to the right, our pirate is a bit of a jester. Again, these are all supposed to be wide-angle shots, so not great.

The images above use the same prompt asking for a full-body view. Three are literal closeups.

Same prompt. Note that sexuality, nudity, violence, and other terms are flagged and not rendered. Also, notice that some of the images include nudity. This is a result of the training data. If I were to ask for, say, the pose on the lower right, the request would be denied. More on this later.

In the block above, I am trying to get the model to face the camera. I am asking for the hat and boots to be in the frame to try to force a full-body shot. The results speak for themselves. One wears a hat; two wear boots. Notice the shift of some images to black & white. This was not a request.

In the block above, I prompted for the pirate to brush her hair. What you see is what I got. Then I asked for tarot cards.

I got some…sort of. I didn’t know strip-tarot was actually a game.

Next, I wanted to see some duelling with swords. These are pirates after all.

This may not turn into the next action blockbuster. Fighting is against the terms and conditions, so I worked around the restrictions the best I could, the results of which you may see above.

Some pirates used guns, right?

Right? I asked for pistols. Close enough.

Since Midjourney wasn’t so keen on wide shots, I opted for some closeups.

This set came out pretty good. It even rendered some pirates in the background a tad out of focus as one might expect. This next set isn’t too shabby either.

And pirates use spyglasses, right?

Sure they do. There’s even a pirate flag of sorts on the lower right.

What happens when you ask for a dash of steampunk? I’m glad you asked.

Save for the bloke at the top right, I don’t suppose you’d have even noticed.

Almost to the end of the pirates. I’m not sure what happened here.

In the block above, Midjourney added a pirate partner and removed the ship. Notice again the nudity. If I ask for this, it will be denied. Moreover, regard this response.

To translate, this is saying that what I prompted was OK, but that the resulting image would violate community guidelines. Why can’t it take corrective actions before rendering? You tell me. Why it doesn’t block the above renders is beyond me – not that I care that they don’t.

This last one used the same prompt except I swapped out the camera and film instruction with the style of Banksy.

I don’t see his style at all, but I came across like Jaquie Sparrow. In the end, you never know quite what you’ll end up with. When you see awesome AI output, it may have taken dozens or hundreds of renders. This is what I wanted to share what might end up on the cutting room floor.

I thought I was going to go through pirates and cowboys, but this is getting long. if you like cowgirls, come back tomorrow. And, no, this is not where this channel is going, but the language of AI is an interest of mine. In a way, this illustrates the insufficiency of language.

Putting the Mid in Midjourney

Posted in technologyTagged , , , , , , , , 1 Comment

I use generative AI often, perhaps daily. I spend most of my attention on textual application, but I use image generations, too—with less than spectacular results. Many of the cover images for the articles I post here are Dall-E renders. Typically, I feed it an article and ask for an apt image. As you can see, results vary and they are rarely stellar because I don’t want to spend time getting them right. Close enough for the government, as they say.

Midjourney produces much better results, but you need to tell it exactly what you want. I can’t simply upload a story and prompt it to figure it out. I’ve been playing with Midjourney for a few hours recently, and I decided to share my horror stories. Although it has rendered some awesome artwork, I want to focus on the other side of the spectrum. Some of this is not safe for work (NSFW), and some isn’t safe for reality more generally. I started with a pirate motif, moved to cowgirls, Samuris and Ninjas, Angels and Demons, and I’m not sure quite what else, but I ended up with Centaurs and Satyrs – or did I?

It seems that Midjourney (at least as of version 6.1) doesn’t know much about centaurs and satyrs, but what it does know is rather revealing. This was my first pass:

Notice, there’s not a centaur in sight, so I slowly trimmed my prompt down. I tried again. I wanted a female centaur, so I kept going.

So, not yet. It even slipped in a male’s face. Clearly, not vibing. Let’s continue.

Trimming a bit further, it seems to understand that centaurs have a connexion to horses. Unfortunately, it understands the classes of humans and horses, but it needs to merge them just so. Let’s keep going. This time, I only entered the word ‘centaur’. Can’t get any easier.

It seems I got an angel riding a horse or a woman riding a pegasus. You decide. A bull – a bit off the mark,. A woman riding a horse with either a horn or a big ear. And somewhat of a statue of a horse. Not great. And I wanted a ‘female centaur’, so let’s try this combination.

Yeah, not so much. I’m not sure what that woman holding bows in each hand is. There’s some type of unicorn or duocorn. I don’t know. Interesting, but off-topic. Another odd unicorn-horse thing. And a statue of a woman riding a horse.

Satyrs

Let’s try satyrs. Surely Midjourney’s just having an off day. On the upside, it seems to be more familiar with these goat hybrids, but not exactly.

What the hell was its training data? Let’s try again.

Not so much. We have a woman dancing with Baphomet or some such. Um, again?

We don’t seem to be going in the right direction. I’m not sure what’s happening. Forging ahead…

On the plus side, I’m starting to see goats.

There’s even a goat lady montage thing that’s cool in its own right, but not exactly what I ordered. Let’s get back to basic with a single-word prompt: Satyr.

Well, -ish. I forgot to prompt for a female satyr.

Ya, well. This is as good as we’re getting. Let’s call it a day, and see how the more humanoid creatures render.