Getting better: With all the recent news revolving around ChatGPT and other large language models, it’s easy to forget that their cousins—AI image generators—are still improving. One may have figured out how to render eyes and hands without making the subject look like something from a nightmare. however, the results still creep some people out.
Earlier this week, research lab Midjourney released a beta for version 5 of its self-named AI-imaging software. According to its announcement via Twitter, the latest version adds higher picture quality, more “diverse” results, a more expansive range of styles, seamless textures, and much more.
Starting today our community can test Midjourney V5. It has much higher image quality, more diverse outputs, wider stylistic range, support for seamless textures, wider aspect ratios, better image prompting, wider dynamic range and more. Let’s explore!
— Midjourney (@midjourney) March 15, 2023
Users have already posted hundreds of stunning results, and feelings about the improvements are mixed. Most are impressed because imaging AI has struggled to produce aspects like shadows, reflections, eyes, and hands. Below is an image we created with OpenAI’s Dall-E as an example of where the machine has trouble.
The composition is somewhat off, and the general feel is cartoonish. The lighting is all wrong. The eyes and hands are badly deformed. The legs are fouled with artifacts, as are the popcorn container and the seat next to the subject. This result is one of four with similar problems to varying degrees.
Version 5 of Midjourney seems to have improved in this respect, at least from the examples others have shared. The results from simple prompts border on the uncanny valley—realistic enough to pass as professional photos in many cases, but still with that odd quality you can’t quite place. While incredibly realistic, many have described the images as creepy.
Midjourney v5 is here! (for real this time, lol)
Here are some side-by-sides of my prompts, v4 vs v5, as well as some new prompts and crowd shots. I’ll add more to this as I experiment.
— Nick St. Pierre (@nickfloats) March 15, 2023
Our own Kishalaya Kundu said, “I’m more afraid than impressed, to be honest,” after viewing a series of nearly flawless Midjourney V5 photos. The fear being that one could fairly easily create a fake image and pass it off as genuine.
Creep factor aside, compared to V4, Midjourney V5 has dramatically improved quality. Graphic designer Julie Wieland has used Midjourney V4 (released last November) for some time and says that version 5 has “incredibly realistic” skin textures. The lighting effects are also much better, including reflections, glare, and shadows. Perhaps most importantly, the AI generates hands and eyes that appear natural most of the time.
ï¿½”ï¿½ MJ tip: shots through a window are finally possible with V5!
I’ve been craving the “My Blueberry Nights”-aesthetic since I first tried out Dalle2 (and it did okay-ish), but v5 is mind-boggling!
ï¿½’ find the prompt in the ALT text of the images #synthography #midjourneyv5 pic.twitter.com/kAOagopucG
— Julie W. Design (@juliewdesign_) March 17, 2023
“Eyes are almost perfect and not wonky anymore,” Wieland told Ars Technica. “Hands are correct most of the time, with five fingers instead of 7-10 on one hand. MJ v5 currently feels to me like finally getting glasses after ignoring bad eyesight for a little bit too long. Suddenly you see everything in 4k; it feels weirdly overwhelming but also amazing.”
1960s street style photo of a young woman, sitting, sailboat, green dior dress, silk green dress, green dress, silk, pearl necklace, tiffany’s pearls, tiffany’s pearl necklace, sunset, ocean, shot on Agfa Vista 200, 4k –ar 16:9
v4 (left) v5 (right) pic.twitter.com/wz7GbI3fvA
— Nick St. Pierre (@nickfloats) March 15, 2023
Midjourney also improved the native resolution from 512x512px to 1024x1024px. The increase aligns it with Dall-E. However, Version 4 could supersample to double the native resolution. It’s not unreasonable to expect V5 to use the same technique to produce 2048×2048 images, but that is for an update further down the road.
The bottom line is MidJourney only hit the AI scene one year ago. Many (not all) of these images flooding Twitter feeds this week are untouched. Previously, Weiland used a combination of techniques to improve Midjourney 4’s visual quality, including “outpainting” with Dall-E and touchups in Photoshop. Version 5 promises less post-generation editing and perhaps photo-perfect images sooner than we can imagine. This prospect is indeed both exciting and frightening.
source : www.techspot.com