A MidJourney Prompting Guide
The best prompting + generative AI gets you photo-realism ... and Deep Fakes.
Midjourney Additive Prompting For Photo-realistic AI Images
Generative AI models can create amazing images when given the right instruction prompts, and we’ll share some tips on how to prompt Midjourney to create great results here.
Photo Credit for the top-line picture goes to Nick St Pierre and Midjourney V5. He used the prompt:
1960s medium-full off-center shot, 35 mm Kodachrome film still, capturing a Japanese woman peacing out and waving down a taxi, wearing a gingham print dress made of silk, blue/white palette, accessorized by sleek pearl earrings, another moody late-night in Tokyo --ar 1:1
Doing the same prompt in Bing image creator gives a result that captures the same essence of the scene, the subject, and the type of film used.
Nick has created highly realistic Midjourney images that look like professional photos of various styles, lighting, moods, and he’s shared his prompting guide how-to on Twitter threads. He starts simple:
A very simple prompt to get started with your own creations [Shot Type] photo of [Subject], shot on [Film Type] For easy comparison, all images in this thread are [street style] photo of [a woman], shot on [Film Type]
From there, he is able to explore different modes within each category. Here's a compilation of his Midjourney threads, showing how he uses his Additive Prompting framework to generate characters/models, design sets, & more. Here are the categories his uses in his framework:
For example, starting with “street style photo of a woman, shot on Afga Vista 400” as the base prompt, he shows how lighting and weather can impact the final result. This is “sunny” versus “snow” weather conditions:
Likewise, how you can use time-of-day to control the lighting and look of the picture. This is sunset’ versus ‘night’:
He also shows various Directional Lighting styles: Front lighting, left or right side-lighting, or back-lighting, and prompts for various lighting patterns including hard lighting vs soft lighting vs diffused lighting. More details on all of this is here.
Combing all these together with specific camera Angles & Shot Types can yield very specific but also stunning and realistic results. Here he compares “foggy night with neon side lighting" and "silhouette lighting with side light.”
With additive prompting, you can get your image generation to be more precise, on target, and impressive. The methods apply on other diffusion-model image generators like Bing image generator as well (although end-result quality is not up to Midjourney v5 levels). For example, using Bing image generation, I got this:
Too Good to Be True … Deep-Fakes
This is where we remind people: This is the worst the AI will ever be in your lifetimes.
MidJourney v5 has gotten very good and will only get better. With the right prompts, these are becoming so photo-realistic and with few artifacts of error that you cannot tell if its AI or a real image. Sometimes there is a ‘tell’, but Nick’s latest tweet is a challenge:
Maybe the shadow of the left earring is a tell, but if you told me this was actually a shot from some movie, I wouldn’t be able to say you were definitely wrong.
You can tell Pope playing basketball is a fake image because 86-year-old Pope Francis doesn’t play basketball, not because of any artifact. Everything looks real, down to the focus on the Pope and de-focus on the fans watching. The dangers of “deep fakes” of prominent individuals increases as these get better.
The mainstream platforms can try to curb abuses. For example, prompting “Pope Francis playing basketball” in Bing Image Creator returns ‘prompt blocked’. But open source stable diffusion AI models can be downloaded and run at will on the cloud or on your own PC, so deep fakes are with us.
Deep Fakes Go Hip-Hop
Generative AI that fakes the human cultural creations is extending into the realm of audio, music and video. The latest news is an AI-generated song that used Drake and Weeknd voices was released and generated over 1 million hits before being taken down: “Fake Drake & The Weeknd Song — Made With AI — Pulled From Streaming After Going Viral”
As an allegedly AI-generated song racks up millions of listens, UMG warns that industry players need to choose "which side of history" they want to be on”
Which side of history is that? The AI has been let out of the Pandora’s Box, and it’s not going back in. Humans will leverage AI to make things, because it’s easier to create than ever. AI is good and will get better.
I’ll leave with the response Sander Deuvestein put out on twitter to this news. The era of AI-generated art - images, songs, and soon movies - has arrived.
“Interesting, but this movement can’t be stopped. Look at the AI Hub Discord Server. Complete access to a community, tutorials, voice models and a working Google Collab environment. As an experiment I’ve created Kanye West singing Flowers. It’s that easy.” - Sander Deuvestein