At the turn of the year, so much has changed, so before I can even get to the still pending topic of „consistent characters,“ I’ll summarize the changes.

For reference my posts Midjourney I and II (older blog system).

For those who, like me, have painstakingly created an Excel list with prompt properties and parameters, stylize parameters, and attributes, rest assured, nothing is lost; you just need to rearrange the columns differently. The prompt structure has changed, and Midjourney has become more receptive to free sentence formations (similar to ChatGPT or DALL-E, for example):

Style+Subject+Setting+Composition+Lighting+Additional Info Photo of an astronaut in a white space suit, helmet visor reflecting stars. Standing on the moon with Earth visible in the starry sky. Astronaut centered, Earth in the background. Bright sunlight with soft moonlight reflections. Moon rocks and small craters nearby.

The parameters remain the same: –ar, –c, –w, –s, –sref (new), –sw, –style raw, –no, –q, –v, –video, –seed, –tile, –r, –stop.

The acclaimed new, yet unavailable, video tool Sora also doesn’t have much more: –scene, –character, –angle, –lighting, –gear, –motion, –pan, –roll, –tilt, –zoom, –speed, –3D, –AGI.

On TikTok’s Boximator Video AI: I can already look forward to it (it will take another 2-3 months).

Midjourney V6, the „slotmachine“, has become good with text, meaning, for example, a sign with the text „Duck“ will definitely bring you a sign with the text „Duck.“

In Midjourney, the keyword now is F.R.A.M.E. (Focus, Resolution, Ambiance, Mood, and Extras). As a framework, for example, an [type of image] of a [subject] with the [frame/composition technique] and a [background description]. Add [type of lighting] in [how it affects the scene] with [additional details] around the subject. So, genre and style, camera angle and shot type, technical details, character description, clothing, setting, time of day, light/shadow, and weather, 35mm/16mm/70mm/8mm/IMAX 70mm/digital formats like 4K, 8K film/additional Midjourney parameters. In America, there are up to 240 hours of free courses through their Workforce Innovation and Opportunity Act. (For Austria: I am not aware of a relevant AMS program.)

Some also refer to S.S.S.C.L.A for a keyword Style, Subject, Setting, Composition, Lighting, Additional Info.

Good, on YouTube, you can learn almost anything. Here are some of my favorites:

https://www.youtube.com/@curiousrefuge, https://www.youtube.com/@delightfuldesign, https://www.youtube.com/@TheoreticallyMedia, https://www.youtube.com/@futurepedia_io (however, his English operates at the speed of Mach 1), https://www.youtube.com/@TokenizedAI, https://www.youtube.com/@AureliusTjin, https://www.youtube.com/@digital_magic, https://www.youtube.com/@cyberjungle)

For cinematic or photorealistic images, you should remember F.O.C.A.L. F for F.R.A.M.E (as mentioned above), O for optimal lighting conditions (including colored light/shadow, hard/soft, gentle, natural, calming, from behind/in front, high contrast, e.g., use fill lighting to soften shadows on the face caused by natural light coming in from the window resulting in a balanced, evenly lit image), C for camera angle, A for aesthetic style (sense of nostalgia, classic portrait, street photography, surreal fantasy, modern minimalism, very trendy gritty noir, etc.), L for lens focus and depth of field (Depth of Field) (Lens focus sharpens a specific part of the image, while depth of field (DoF) dictates the extent of sharpness from front to back. Manipulating focus and DoF can isolate subjects, create dreamy backgrounds in portraits, or achieve crisp clarity in street scenes).

Regarding the new Midjourney parameter, –sref = style referencing.

The „Style References“ option works similarly to image prompts where you use an image as a reference for your prompt. However, while an image prompt focuses more on copying the composition of the reference image, style referencing emphasizes transferring the overall aesthetics.

Example:

For the painting – based on a tulip-lilac image of my grandfather – available here (https://gabriele2500.com/products/thin-canvas-2?variant=48010089234760), did I use –sref? No, some parameters and –style raw (as similar as possible in terms of composition without Midjourney aesthetics or any style like van Gogh).

Here’s a simple example with another image of my grandfather (without removing the background and his logo beforehand, so quite brutal – I can’t influence the background much, it would be better to upload with a transparent background).

Prompt 1: a photorealistic painting of a Japanese quince twig, background skies blue –sref https://s.mj.run/38h9UJPPuxs –ar 2:3 –v 6.0

There’s absolutely nothing to complain about; it clearly reflects the aesthetic of grandfather Lichtenstrasser.

Prompt 2: an editorial image of a living room, blue denim-coated table and high-end seats + minimal decoration + metallic chandelier –ar 2:3 –sref https://s.mj.run/38h9UJPPuxs –v 6.0

From a copyright perspective, the generated image is in the style of grandfather Lichtenstrasser.

Prompt 3:

More or less the same prompt as in prompt 2 but with pillows of the link provided –sref https://s.mj.run/38h9UJPPuxs

Magically, Midjourney has also incorporated the details of my actual living room windows.

Prompt 4:

Here, I used the following prompt and utilized a combination of sref and Imagelink (the reference image for this was the lion from the title image of this article).

hyperrealistic lions head with 3D gradients Isolated on a white background, 35mm, Kodak film –sref https://s.mj.run/38h9UJPPuxs https://lion –v 6.0 –s 130 –ar 3:2

Yes, I can see, in the style of grandfather.

Prompt 5:

Here, I used some image of a bar as a second reference.

a painting of a Japanese quince in the bar https://bar –sref https://s.mj.run/38h9UJPPuxs –ar 2:3 –v 6.0

Amazing, amazing, amazing. Please add it to my house immediately. I might paint the mural myself.

Additionally, you can influence the –sref references with other parameters:

parameter —sw {value from 0-1000}

You can also provide multiple sref URLs, two styles from my grandfather and/or mix them.

–sref urlA urlB urlC

And you can assign weights. The second reference has a stronger weight.

::{weight}

--sref https://something ::1.5 <https:// https://s.mj.run/38h9UJPPuxs>

What other news is there?

For those who want to stay up-to-date on AI, I recommend once again the newsletter from AI Fire.

🔥 🔥 🔥 EMO – Academic Research, Institute for Intelligent Computing, Alibaba Group

🤩 The Generation Z is the „old soul“ generation (there’s a lot that comes to my mind about this, but let’s leave it at that for now). They embrace dinner at 5:00 PM and bedtime at 8:00 PM, and the comeback of jazz suits me very well. I also remember recently reading that „surge pricing“ in restaurants has suddenly become trendy, and an Austrian Twitter user posted a desperate picture a few days ago of his sandwich (Semmel) with two slices of ham for 12 euros (he must have caught such a surge pricing moment).

🧓 https://www.sueddeutsche.de/wissen/archaeologie-fund-schlangen-bayern-awaren-ungarn-guertelschnallen-1.6322951

🔮 material revolution (machine intelligence and nanotechnology, kintsugi (upcycling)), boundless multidimensional data (quantum computing, blockchains, loT (like smart homes and industry), edge computing, automation, 5G, 6G, generative AI), technological vulnerabilities will become more complex, energy boundaries (water eater AI!, solar energy, geothermal energy), saving ecosystems (maintaining biodiversity while meeting basic human needs, Bio Trade), borderless world -fluid economies (Increasingly unmediated transactions in finance, health, education, trade, services, and even space are leading to the blurring of jurisdictional boundaries, changing liabilities, and increased numbers of cross-border communities. Advances in communications, computing, and advanced machine intelligence will accelerate a borderless world that will change the way we work, live, and connect.), digital realities, living with autonomous robots and automation, future humanity (brain–computer interfaces (BCIs). New definitions of self-esteem, autonomy, and stability will bring forth new ideas about parenting, care, love, belonging, inclusion, and community.), advanced health and nutrition

📚 Anything by Tim Marshall (Anyway, for people interested in politics.): z.B. Prisoners of Geography, The Future of Geography (Prisoners of Geopgraphy: he updates every now and then the very same book. At Zurich Airport I have seen over the years at least three updates (and interestings ones).

Never split the difference (by legend Chris Voss und Tahl Raz). The Swan-Trick is amazing, and courses would certainly be recommended in German as well. However, and he and his team admit this, women cannot implement it 1:1, and in German, it is also difficult. Fantastic glimpses into his professional life, and I love him for his probably inherent humor and the phrase „and how am I supposed to know … dog.“

Million Dollar Weekend by Noah Kagan. I haven’t read the book yet, but I know the story about why and how difficult the decision was to choose the green cover. The book cover decision definitely took longer than a weekend 😉

👔 Microsoft’s AI principles promote innovation and competition and “eat” the only Open Source Hope of Europe (France’s Mistral AI).

🐦Instead of showing the bird Sendbird builds a custom GPT on your website and mobile apps to automate engagement, marketing, sales, and support with conversational AI.

🎰 Markov Chains Monte Carlo:  In generative AI, they serve as a foundation for generating sequences of data points based on the probabilities of transitioning between states. They have however pros and cons. And are there alternatives?

As we know from my blog Midjourney II (see a bit above) 😉, there are generative and discriminative models. What exactly is the „language“ (=algorithm) difference, and why is no one talking about it?

Andrey Markov (Russian), who did not know Stanislaw Ulam (Polish-American mathematician and nuclear physicist involved in the Manhattan Project, who had an idea while playing solitaire), are the namesakes. Due to the sensitive nature of the Manhattan Project, Ulam needed a code name, and the Greek physicist Metropolis (no joke, for real) came up with the name Monte Carlo (apologies, gambling in Las Vegas had just been legalized shortly before). The MCMC algorithm is a deterministic function of the simple random number generator (RNG). And with every spin the slot machine uses RNG. In the broadest sense, one could indeed use the technical term „Slot Machine“ for AI, without any negative connotation.

GIGO: garbage in, garbage out. Second rule of MCMC

to be continued …