Woman in pink kimono fights a circle of samurai — AI art by @user_google_961350

CaylaCatz

Yeah, Nano can be very restrictive. Google doesn't want people to protest them for using bad words or making sexy or violent images. I think outside the USA, companies may not have to worry as much about being sued. So you might want to learn a little styling so you get NanoBanana-like results on other models when you have to switch. Also I haven't said it but you are already learnign the harder part of prompting -- making sure the people are in good poses, holding weapons right, in good positions, all looking in the correct directions. It's a lot harder to make a working prompt with mulitple people. Adding action to that is another layer. So you are also learning the harder part of prompting first. Just add the prompting for style part when you're ready for it. Every model can make the style you want if you prompt it. They all have millions of images they were trained from. Your prompt helps them to access the right type of images to analyze for patterns and create a new image for you

CaylaCatz

It's ok to have a prompt with a few words for style. Some models are good with that. I'm a fan of doing that actually. Sometimes you want more control tho. You know how GPT 1.5 low has a very strong signature style on BP? DavidP uses a prompt to overide that style to make the bright painterly style he prefers. He did a blog post showing how he changed his prompt step by step from GPT signature look to his signature look. You might want to see the difference prompting makes to override a style in case you want to do that https://budgetpixel.com/blog/prompt-development-iterative-increments . But if you found a model you like, as you use it a lot, you'll know it's nuances and you'll be able to make your fav model do things that other people don't know it can do. Also scenes like this are great start images for making videos.

CaylaCatz

Well, their governments might be restrictive about what's seen in their own country but may not care what's seen in other countries. It's not their people pushing for the restrictions. In some Western countries tho, the restrictions come from people who want to push their belief system onto other people or feel that limiting other people's activities is the only way to protect their children since they're busy at work or protesting and can't spend time with them. So they make companies afraid by suing them or creating bad publicity. Trying to control the behavior of everybody else in the world isn't going to work tho. Eventually someone will make an AI app that monitors incoming data on the home's wifi. That will probably be a better solution when that happens.

CaylaCatz

I don't know if this will help. It was old so I asked copilot to update it for now. It seems to work, better than my chunk before actually : "sweat, dust, and fabric fibers visible, natural imperfections, grounded physicality, hard‑edged contrast, practical lighting, subtle lens breathing, crisp highlights on steel" . My old prompt used the name of a camera, a lens type and the camera settings. Copilot said that gets treated as a style rather than instructions. So if you notice in mags, that people like to take martial arts / sword fighting scenes with a particular camera and lens, try putting that in. I had Hasselblad in y prompt and it's more known for stills like museum artifacts, not really a camera to photograph people or sports.

user_google_961350

I do routinely include words like "photorealistic" or "cinematic" in the prompt, but the different models seem to have different standards for those. SeeDream, for example, often sacrifice realism for visual effects, I feel. I don't use Minimax that much--for the same reason you mentioned. It tends to give me surprises, and I don't like surprises. 🤣 It seems to have a more limited database to draw images from, and as a result a lot of the objects turn out close enough but not quite right.

user_google_961350

Thank you very much! I have never tried this much details in the stylistic instruction in a prompt; usually it's just a four- or five-word phrase at the end. I did notice, though, that many people here give AI very specific and detailed instruction on art styles, often referencing particular artists or, as you mentioned, equipment. I've gotta give that a try--rather than simply telling AI "make it look like a real photo"--and see how it comes out.

CaylaCatz

Have you tried embedding words like hyperrealism, photorealism? I think if you try out some words with that kind of flavor, you'll find Flux 2 Pro and Grok does pretty well too. I find Minimax goes very realistic but it doesn't follow a prompt very well and tends to have imperfections. But I feel like MInimax might have been trained on more cinema type images, maybe independent films, because it gives that flavor.

user_google_961350

Very interesting and enlightening stuff. I have to admit that I haven't studied these AI models enough to develop a good understanding of them yet. Right now Nano Banana is definitely my preference, but I end up using SeeDream, WAN, and Z Image more because they are less restrictive on what you can do. Too many rejections from Nano Banana!

user_google_961350

It's rather ironic that SeeDream, WAN, and Z Image, created in the most restrictive cyber realm (China), can be used without restriction here on BP and on many other platforms. I guess since those are open-source, ByteDance and Alibaba can simply shrug and say innocently: "Don't look at me...I ain't got nuthin' to do with this."

CaylaCatz

I just think it depends on the image, the art style and how much you want going on in the picture. So I jump from model to model a lot. But I don't do reality much so that maybe why I don't have a set favorite. Reality is a whole different ballgame than fantasy. Nano did great here.