This content was created by LARRUZ. This content was uploaded by LARRUZ. This is a YouTube tutorial/demo (title: "MidJourney GAME Changing NEW FEATURE - /describe (Reverse Engineer PROMPTs from IMAGES)") published April 4, 2023, in which a single presenter walks through MidJourney's new /describe feature that converts images into generation prompts. The presenter tests the feature on multiple example images — a cat in a coat in the rain, a Joker image, the Eiffel Tower, human faces, and headphones with colorful smoke — comments on generated prompts/styles, runs generations, notes hits and misses (style/face fidelity, aspect ratios), highlights remix/edit options and a Stable Diffusion/CLIP analogy, and gives an overall positive first-impression/usefulness assessment.
Chapters
There's a new awesome feature in midjourney called Describe. It lets you turn an image into prompts. I'm going to test this new feature with these images of different textures and complexities and shapes. There are multiple different objects which would be nice to see if it can actually recreate these. In order to access this new feature and simply type in slash describe. Then it gives you an option of adding or uploading an image. So let's say we will first upload this image. It's an image of a cat wearing a coat in the rain and there is a city in the background. There are cars on the road. Let's see what it can come up with in terms of the description. So let's simply run this. Okay, so here are different description that it came up with. So let's look at a couple of these. A cat standing in the city in the rain in the style of Art Groom. It actually gave a reference as well. So that's nice. Handsome, realistic, hyper detailed portrait. Contrasting lights and drags. Okay, actually let's look at this style. Let me click on this. And it's a style of Stanley Lou. Okay, this is another one. The city life of a cat's cat. Interesting in the style of realistic and hyper detail renderings. Cat scat. I don't know what it means. Okay, let me run a couple of these. So I'm going to run the first one and potentially the third one. Okay, so here are the results of our first run and it's actually not bad at all because if you look at these images, they are really in the style of our original image. So I'm really impressed because there is rain in it and there's a the cat is wearing coat and then there is a city in the background. This is pretty good, these ones. Okay, so for example, the second one, it's a normal cat, not really a character cat. But the other one looks pretty good. I think it works. Let's try it on the Joker image and let's see if it can actually capture the style of the image. So I think it will definitely recognize the Joker but I'm not sure about the style and texture of the image. Okay, so here are different in four different prompts that came up with the Joker comic art wallpaper in the style of Magali. But it actually pretty good because it gives. Although I think the style is very different, it does give you a reference style as well. So that's pretty nice. And actually a lot of artists are going to be happy with it now. Okay, let's try to create an image using this and let's look at some other descriptions. So the joker is in the black and purple light in the style of light green and dark crimson precision painting. Okay, let's try this one as well. Okay, so this is our original image and this generation definitely missed the mark in terms of like hairstyle and facial expressions. But I think these ones are pretty nice, especially the first one. The style of the image is pretty consistent with what we had here. So it seems to be working like you need to experiment with it a little bit, but I think it's something very useful. Now there is an a tool for stable diffusion based images. So it's called Clip. I have a video on it, so watch that. But having it for Mid Journey itself is awesome. Okay, let's experiment with a more abstract art. So here I have this image of Eiffel Tower. Let's see what it thinks of it. Okay, so the Eiffel Tower in colorful color sprayed in the black paper in the style of flowing form. National Geographic photo. Eiffel Tower is colorfully drawn on the black. Eiffel Tower is drawn in the paint or black back background. All right, let's try a couple of these and let's see if we can actually get a similar image or not. This is going to be really awesome because online sometimes you find these images or on Reddit people always complain that people share their images but not the prompts. So I'm assuming no, they don't need the prompts. They can recreate these images using prompts from Mid Journey itself. All right, I think in this case this is our original image and here are the first generations and this is with the second prompt that we tested. So maybe here it really missed the mark because it's not the same style. It's Eiffel Tower, but the style is completely different. So I think it could be like a hit and miss. But if you think about this image, it is really a hard one to describe. So maybe that's why it couldn't really come up with a better prompt. Okay, let's look at human faces. I'm going to use this image as an example. Right. It's still an artistic depiction, so let's see what it thinks of it. Okay, so there are multiple prompts, but I actually realized just one thing. If you click on this, these numbers 1, 2, 3, then it kind of gives you the remix mode and you can actually make some changes if you want. So this is pretty neat actually. This is awesome. So let's do the first and the last one and let's see what the results look like. But this is a great feature, so. So it not only lets you generate the image with this prompt, but if you want to modify the prompt, you now have the ability to do that pretty good. Okay, so here are our results. This is nice. This is nice. But I think the color scheme is a little bit off compared to what we had with that. And since it's a face, so you can't really expect it to be a very similar face. But I think overall, not too bad at all. All right, so the last test I'm going to do is I think going to be a very interesting one. So I personally really like this image. There is an image of headphones and then small colorful smoke is coming out of the earpiece. So let's see what type of prompts it can come up with. Okay, so if you look at it, a headphone with color smoke coming out of it in the style of a dark orange and dark black. That is awesome. So this is dark orange. Dark orange and dark black or bluish, basically. Headphones coming. This is pretty nice. The aspect ratios are a bit weird, I think, in this case, and this is the first time I actually saw aspect ratios. So let me generate the first one and the third one and let's look at the results. Okay. Some pretty neat results. Although, like in the first two images, actually in the fourth image as well, it put two headphones, but the smoke is coming out of them, so that's pretty nice. In the second set, there is just one headphone in each of the image. Although I still look like the style of the original image. But these are actually not bad at all. These are pretty awesome. Now, what do I think of the feature? It's actually pretty awesome feature. I really like it. And this was just the first impressions, but I think this is going to be a very useful feature for a lot of people. People will come up with some very interesting use cases. Well, thanks for watching. See you in the next one.
