세부적인 프롬프트 기법을 통한 이미지 생성 최적화

### 효과적인 프롬프트 기법을 통한 이미지 생성 최적화
* 🏡 "새 집으로 이사하는 바람에 이 동영상의 설정이 이상적이지 않아서 방해가 된 점 사과드립니다."
* 🖼 오늘날에는 블록 프롬프트에 중점을 두어 단순한 쉼표로 구분된 문구보다는 **고품질 이미지를 얻기 위한 구체적이고 상세한 프롬프트**의 중요성을 강조하고 있습니다.
* 🔍 이 동영상은 **클립 텍스트와 코드가 포함된 워크플로우**를 사용하여 프롬프트를 최적화하는 방법을 안내하여 생성된 이미지를 크게 개선할 수 있습니다.
* 👩 예를 들어 **'25세 여성'**과 같은 세부 정보를 지정하면 생성된 인물 사진의 사실성과 디테일이 향상되며, **인종을 포함**하면 결과물을 더욱 세분화할 수 있습니다.
* 💡 '**차가운 눈'이나 '윤기 나는 피부'**와 같은 특징을 자세히 설명하는 등 묘사적인 표현을 사용하면 더 정확하고 시각적으로 매력적인 결과를 얻을 수 있습니다.
* 📊 또한 이 동영상에서는 **ChatGPT 또는 Copilot과 같은 고급 프롬프트 생성 도구**를 사용하면 더 나은 이미지 생성으로 이어지는 상세한 프롬프트를 만드는 데 도움이 될 수 있음을 강조합니다.
* 🌄 또한 풍경 이미지에도 동일한 원칙이 적용되며, **설명이 더 많은 프롬프트가 더 풍부하고 다양한 결과물**로 이어질 수 있습니다.
* 📈 궁극적으로 플럭스 모델에서 고품질 이미지 결과를 얻으려면 **프롬프트의 상세한 설명**이 중요하므로 사용자가 **프롬프트 기술을 향상**시키는 것이 필수적이라는 것이 핵심 요점입니다.

효과적인 프롬프트는 AI로 생성된 아트워크의 이미지 품질을 크게 향상시킬 수 있습니다.

Hello everyone, welcome back to the channel. So first, before we begin with today's topic, I would like to address that I've been moving to a new house. Actually, I'm in the new house right now, so you can see that the setup is not ideal. It's not like the previous videos, so I apologize for that. Also, I'm missing a door there. There's no door for now. It's going to be like this, so if it is too distracting, let me know in the comments down below, and I'll probably put the eain just to hide that section there. All right, so move on to today's topic. I'm going to talk about blocks prompting. A lot of people say that the way that flux is, it's quite hard to get the actual image because of the prompt. You need to have a very specific prompt. You may have seen reddits talking about how you can output comma-separated prompts like one girl anime blonde, and then all of those are separated by commas. Instead, you need actual text. You need complete phrases, sentences just to

get a good prompt. So, as you can see on the screen here, I have a prompt. If you'd like to get these types of images, this is the video for you. Also, if your images are like this, where you have a little bit of deformities, again, this is going to be a video for you. Now, if you're already an expert in prompting, or you are a really good prompt engineer and you're able to get images like this, then I don't think you are going to get any benefit watching this video. You can easily skip it. I've already made a video on prompting before. At that time, I talked a little bit in more detail on how Comi was working with those texts, what was happening in the conditioning node. In today's video, I'm not going to go into that much detail. Instead, we are just going to look at the workflow, mainly the clip text and code, and I'm going to talk about how you can optimize your prompt in order to get images that look like these ones for the workflow. I'm using

the basic workflow. We have the load diffusion model, the Dual clip loader, load VA. Those are pretty standard. If you watched my previous video, you know that I've been using this workflow for quite a while. I'm also using the clip text en code for flux. That is because it provides two fields for the prompt. The top one is for the clip L, and then the bottom one is going to be for this T5 XXL one. This FBA does not matter if you have the fp16; you can put that there. All right, so if you're getting images like these, which is pretty okay, except for this section here, it's mainly because you are typing your prop like this: a portrait of a woman. And right now, you're telling the model to pick from everything that was trained on women and just make a portrait out of it. So, if you really want to get those high-quality images like this one, where you can see the skin, the eyes, there's a lot of detail in it, or if you really want to go

above you can get these types of photographic images as well, and it's all by The Prompt. I'm not using any Laura. I'm not using any fine-tune version of the flux model. It is the flux one Dev, the fb8 version, and the only thing that I'm changing is the pro. Let's start with this one. I have my seed as fixed. I will also be sharing the workflows, so if you can just drag the images inside of conf, you want if you need any of the settings from the workflow here. So just putting it as a portrait of a woman, we get this type of image. Now, if you add a little bit more detail to that woman, let's say you see a 25-year-old woman. And this time when you click on Q promp, this is the image that we get, and the style is basically coming from the top one here. I have the anime style drawing. That's basically where the style is coming. Let's actually remove the style, and we can just copy everything there, and let's

do one more now. This is pretty good. You can see that the 25-year-old is coming out. We have the stylish hair, we have the clothes there, we have that v-neck also showing. So if you really want to get a good image, adding the age of your subject will influence that final image. All right, so now we have a realistic. It probably was trying to generate a necklace, but there were nothing in that necklace there. So I think just generating a new seed will probably fix it, or we can add to the prompt. But we can see in the face now we have a little bit more detail in it, but we can do better. So we can go back, and right now we have a 25-year-old woman. We can also add the ethnicity to this section here. Let's say we do a 25-year-old. I think this will have to go behind the old, so it's going to be like this: 25-year-old African woman. Let's give this one a go. Oh, by the way, there's a window.

here and we don't have curtains right now, so that's why you're saying the webcam a little bit washed out. So now we can see that we have a person with a dark skin tone, also have K ears, which is usually associated with African women. You can also do American-African, you can do Japanese, you can do Chinese, you can do all of that in this here, just replace this section here. Now aside from that, you can also describe the features that you want to see. I already have one here. This is not the right one. This one here. So in this one, I went a little bit more. I have 25 years old, I have cold eyes, and this is giving the blue in the eye. Usually when you say hot, it's associated with red, and then when you say cold, it's associated with blue in photography at least. We have glossy skin, no makeup, we have skin pose, we have texture clearly visible in focus, and that is what we get here. Let's try to do from a different location, so this time I've changed

African to American. Maybe it's having difficulty generating the blue eyes for African women. So even with this, it's quite short. Like you can't come up with this PRP. We are still getting really good images. That's why the flux model is really good at generating these close-up portraits of people. It's also good at generating landscapes. I will show you that in the next section. Okay, so I'm not going to call this blue, but you get the point there. You can keep on adding. Let me show you the example for that monochrome photography. And I think I need to copy this prop into a notepad here. This is the prop. So actually, yeah, the scroll bar is still there. So it's still scrolling at the top. Have a description of what I'm looking for. And then we have this, of course, the monochrome is coming from this monochromatic frame. And then we have a description for the complexion, skin texture, the eyes, the brows, cheeks, bone, jawline, and neck. And then we have the hair expression at the end. Now, of course,

I did not come up with this PRP myself. I just used LLM to generate the prompt for me. So right now on this machine, I do not have CHBT. If you have CHBT, you can use that. Otherwise, you can use Copilot. I have Copilot on this one, so it's basically the same thing. This CHBD4 and my prompt was to write a description of a photo portrait of a woman in as much detail as possible, focusing on skin textures and features. I wanted that, and of course, it gave me all of this. You can also modify this prompt. You can use CHBT. I think that's going to be the best. But as you can see, even this long, like almost one-page length PRP worked, and I did not notice any difference in the time. I think I still have everything here. This is number two, yes. So it's basically the third generation I've done this. This is where I've started. This is the first one where it generated, basically loaded the model, and you can see that on an Nvidia RTX 4070 TI. It

took 33 seconds to generate the first image. This is the second one, again same time, and then the third time. It probably took a little bit longer because of the clip text. Had to encode that massive paragraph, like basically full-page text. So, it's 37. When you change the text back to something simple, it goes back to 33, 34. And you can see it's pretty consistent. Now next, let's go back to this last image that we generated. And now I'm going to add elements in the background as well. So, this was just the subject. You can describe the foreground if you have any. Like if your subject is the middle ground, and then you have foreground, middle ground, and then the background, you can describe that for too. Maybe you can say a go-to frame. So, this would be an example of describing the foreground if you have bushes. If you're trying to do a more complex scene like bushes in the foreground, and then we have the character in the middle ground, and then

we have some background can do it like this so foreground sub and then you can add the background to now if you don't want to come up with these text one example I've already given you is to use existing llm so those would be those online features that we have ch GPT co-pilot otherwise you can also install llms inside of confy UI you can go into the manager custom nodes the top here just search for llm and there are a couple of llm that you can use I think we also have Ama now there are a couple of nodes here which says conf UI AMA maybe this would work so let's give this one a try basically it says describe the image shortly in this it's using the L lava to describe the image and you can get a prop like this I think it should if you change the model you should be able to generate from just Tex as well but there are those examples you'll of course need the RAM and vram depending on which model you are using all right.

So we can see that now we have that frame. We have the portrait. We also have the lighting, not lightning, sorry, lighting, um, studio lighting shining from the back, and it's clearly here. So if you're able to describe your subject, the foreground, background, the lighting condition, if you can go and describe the clothes that they are wearing, whether they are wearing makeup or no makeup, all of that is going to affect your final generation, but that's mainly for people. Let's focus on landscape. I've just changed the clip text and code to say a beautiful landscape. So you will see a similar pattern here that if your prop is generic, you are going to get one of the generic images from Flux. Okay, so this almost looks like Windows XP wallpaper, and yeah, almost similar ones, right? But now let's actually try to use the copilot. And so again, even though I've said a single paragraph, it just feels like um, copilot is not able to follow those props to the letter as ChatGPT is able to do it. But anyways, copying this into the prompt, let's

generate and see what we get. And just in case you're curious what the problem is, this is the title. Came up with the title, and then we have different, I guess, different sides, different topics of it, invitation as well, okay? And this is the image that we get. And you can, of course, add the style that you want to it. So if you want fantasy anime, you can do all of that. Now there's, of course, luras. You can use luras that are n is specific to just style your image, so you can use that as well. And this technique here of just getting an llm to generate your prop and then just copy and paste it there will work on any web UI. So if you have fukas, you have stable diffusion Forge, I believe stable diffusion Forge can also do flux. Now there was an update where it's able to generate images using flux. Last time I checked, I believe was last week. It cannot do control net IP adapter. I did not try the luras, but it

cannot do those yet, but it can generate images just from a text prompt. So, you can use that inside there as well, and people complaining that you need to come up with a really good prompt in order to generate images. It's not really true because I can say I think I will have to change the resolution for this text because I'm trying to generate something on the left and something on the right, right? And as you can see, it's simple. The only descriptive part here is that I have on the left here and then on the right side here. That's it. I'm pretty sure it should be able to generate something like this. There you have it. We have Batman on the left side and then Superman is on the right side. It added this blur. This is in most of the images I've seen, like even the not these ones, but you can see like we have blurring back here. We have here as well. We have this blood. If you don't want that, you can prompt it to have something

in the background that is still in the focus. Okay, so I hope this video showed you that you don't really need a good prompt as long as you're able to describe what you need. And like I mentioned, this video is nothing new. It's just that when using flux, everything that you've learned for SD 1.5, SD 2.0, 2.1, and SD XL comes into play more here. The flux clips really like this descriptive text, so if you can come up with those descriptive texts on your own, you can paste it there. You are going to get really good, amazing images. All right, thank you for watching. If you like this video, like it. You'll probably get a popup to watch another one if you want to. And if you have not subscribed, click on the Subscribe button. It helps the channel. Thank you to all. I will see you in the next one.

Improve your flux image generation with prompts, explained!

Code Crafters Corner

더 나은 결과를 위한 이미지 프롬프트 최적화

화자는 다른 설정의 새 집에 있으며 블록 프롬프트에 대한 주제로 넘어가기 전에 산만함을 해결합니다.

더 나은 이미지 출력을 얻으려면 쉼표로 구분된 키워드가 아닌 완전한 구와 문장을 프롬프트에 사용하는 것이 중요하다는 점을 강조합니다.

이 동영상에서는 클립 텍스트 및 코드 워크플로우를 사용하여 프롬프트를 최적화하는 방법에 중점을 두어 시청자가 보다 디테일이 살아있는 고품질 이미지를 얻을 수 있도록 돕습니다.

프롬프트를 사용하여 사진 이미지 생성

이미지는 Flux 모델의 개발 버전을 사용하여 생성되었으며, 프로 설정만 변경되었습니다.

프롬프트에 연령 및 인종과 같은 세부 정보를 추가하면 최종 이미지에 영향을 미쳐 더욱 사실적인 묘사가 가능합니다.

눈 색깔이나 피부 질감과 같은 특정 특징을 설명하면 생성된 이미지의 정확도가 더욱 향상됩니다.

AI 모델은 아프리카 여성의 파란 눈과 같은 특정 특징을 생성하는 데 어려움이 있지만 클로즈업 인물과 풍경을 생성하는 데 효과적입니다.

이 과정은 언어 모델에 의해 생성된 상세한 프롬프트를 사용하여 원하는 사진 인물 사진을 설명하는 방식으로 이루어집니다.

사용자는 피부 질감과 특징에 대한 포괄적인 설명을 목표로 프롬프트 생성 및 수정을 위해 Copilot과 같은 도구에 액세스할 수 있습니다.

텍스트 설명이 포함된 이미지 생성

이미지 생성에는 33~37초가 소요되었으며, 복잡한 텍스트의 경우 시간이 더 오래 걸렸습니다.

간단한 텍스트를 사용할 때는 일관된 시간이 관찰되었습니다.

이미지 설명에 전경 및 배경 요소를 추가하면 더 복잡한 장면을 만들 수 있으며, 기존 LLM을 사용하거나 UI에 설치하면 이미지 설명에 도움이 될 수 있습니다.

피사체, 전경, 배경, 조명에 대한 설명은 특히 인물 이미지를 생성하는 데 매우 중요합니다.

일반적인 소품을 사용하면 비슷하고 일반적인 이미지가 생성되는 반면, Copilot과 같은 AI 도구를 사용하면 이미지 생성을 위한 구체적인 프롬프트를 생성하는 데 도움이 될 수 있습니다.

언어 모델을 사용하여 소품을 생성한 다음 이미지 생성 도구에 복사하는 기법은 다양한 웹 UI에 적용할 수 있습니다.

텍스트 프롬프트에서 이미지 생성

이미지 생성 도구는 간단한 텍스트 프롬프트로 시각적 요소를 만들어 완벽한 프롬프트가 필요하다는 고정관념에 도전할 수 있습니다.

위치, 초점 등 원하는 요소를 설명하면 배트맨과 슈퍼맨의 이미지를 생성하는 것과 같은 인상적인 결과를 얻을 수 있습니다.

설명 텍스트를 사용하면 도구의 효율성이 향상되어 멋진 이미지를 더 쉽게 만들 수 있습니다.