Hello, last week, there were really a lot of things happening. SD 3 is now open as of the 12th. The expectations for the local model were quite high, and I was also one of them. However, I couldn't record the video right away because there were office moves and I felt I needed to look into the content a bit before talking about it, so I couldn't record it first. I wanted to deliver it quickly, but now that things are somewhat organized, I think there are some points to convey, so I decided to record a short video. Today, I will briefly talk about installing SD3 and its current situation. Let's take a look at it right away. It would be nice to take a quick look. Stability AI announced the SD 3. In fact, I had the opportunity to experience the API SD3 model a while ago.

Instead of being able to install it locally, this part is about SD3 Medium. This model can be received by logging into Hugging Face. Once you enter here, although it's a model, there are EL numbers. After a while, let's take another look. Yes, when you go to the Ready Cira space, there are many disappointed looks regarding SD3. SD1.5, on the other hand, shows a clear appearance, but especially SD3 seems to be showing the grassy ground. As you scroll down, yes, things like this are seen, right? It really doesn't look like a human figure. In fact, SD3 came out as a pretty decent model, but the results are really devastating, with many negative aspects like this. It's really serious. This is not the appearance of any human, and it seems like it came out like this because of too much censorship. In fact, I tried about 100 times with SD3, but the result is almost real.

There is nothing I like, but they really pick out things like objects or product images well. They are doing a good job of picking them out similar to actual photos, so I don't think the model was made poorly. Therefore, a lot of disappointment is being shown towards SD 3. It's interesting. Anyway, things related to SD 3 were continuously coming out a while ago, but now it feels like they are fading away. I also had high expectations for the collapse of the wall of SD 3. Since my expectations have now collapsed, the disappointment has grown even more. Thanks to the company Stability, we have been able to encounter AI images and use them for free, enjoying many benefits like this. However, Stability is actually recording losses while going through the process of monetizing these models.

Sometimes, I also think of some problems that have arisen. So, I will briefly look at how we can now install and use 3, and then move on. Now, it is possible through the Compi UI. I have slightly modified the workflow, which I will share separately. Originally, when you log in here, you can access the workflow and use everything here. You can also check the demo images, but they are not very satisfactory. In cases like this, the existing Clip or Clipwa text encoder models should have been used, but now, with the T5 model, which has a larger capacity, transitioning to tasks like mid-transfer or ideogram improves text recognition. Similarly, now with SD 3, text implementation like this is possible through the T5 model, making it easier to see that it is achievable. If you look here, you can see all of this.

You need to receive it, and then you can use the SD3 medium model. In this case, the capacity is large below this. Since the encoder model is included, you can use it after receiving it. If you use this, more nodes will be added. So, when looking at the company, I inserted the basic model S SD3 medium model as a checkpoint model, and then I needed a clean model. By activating the Triple Clip Loader, I added Clip G, Clip L, and T5 models to create like this. I activated the SD3 Clip Text Workflow node for the three and created the image. Now, several models of Laura have been released because the SD3 model has been released.

It's about connecting with the insert, and if we remove things like this, we don't see a big difference in our basic workflow. Now, because the Clip model has changed, we can see that the Clip text encoding here has changed a bit, and with positive and negative prompts, now through K-sampler, we go through the VA decode process to produce these images. As I mentioned earlier, images of nature, objects, and things like that seem to be extracted quite well. However, although the portraits are not bad, they feel much lower in quality. When we look at cases like this, faces often collapse a lot, and because some restrictions are placed, that's the case. In this case, I'm not really satisfied with the text recognition again. Sometimes it works well, sometimes it doesn't. However, in this way, when it comes to images of cars, nature images, and things like that, it's really practical.

I can see that it expresses closely. Even with such a good model, the reason for causing so much disappointment is due to the flexible nature of certain individuals. By pursuing this, it didn't turn out well. On the contrary, the image of hiding hands or not showing feet doesn't look so bad. It expresses well, close to the watercolor or realism. Nevertheless, the tangled phenomena behind the grass are really unpleasant to see, causing significant disappointment. In fact, SD 3 doesn't seem to be significantly bad, but there are many issues because of such things. However, there is a bigger problem. I will talk about it again. I brought the sequence that shows this well. Now, we SD users, including myself, are part of this.

I'm talking about having this for free, and then at some point, everything that is created will all be mine. This is how the Stability AI company is now arguing about licenses. So, what it means is, if I were to use this commercially, then I would naturally have to pay for the license costs. If I use this image to train someone to create another SD 1.5 or SD Excel model, for example, then I would have to pay the license fees for that as well. And regarding these license issues, something continuously derives from it. If models created by licenses do not continue to pay license fees, they will eventually stop being used. I really, really like this and will continue to do so.

Using it and wanting to use it commercially, but now you can't use it anymore, so it becomes a big problem. If you continue down here, you can check this structure and confirm some contents related to commercial usage. Yes, when you go in here, you can see that if you pay the cost here, you can use it with a license for commercial purposes. However, there are also conditions attached, so a company with an annual income of less than $1 million or similar can use it. So, small businesses may think it's okay to use it for around $20 a month, but in fact, the number of uses is limited. There are traps that need to be read in detail, making users more angry, right? Where could they be?

If you keep going straight, you'll find it here. There is a limit of 6,000 images here, so not only total revenue but also the number of images is restricted. Due to this limitation, users cannot see many models or images produced by this company. Therefore, unless users continue to pay additional licensing fees, they will no longer be able to use it. It seems very complicated because the criteria are too complex, so the situation is not accurately established. Reflecting this situation, today, in the case of CB AI, the sd3 model has been somewhat released, but now it has been greatly reduced. So, unless the situation is accurately established, it seems like they do not want to create a situation where the sd3 model is not used or problematic.

It's a bit unfortunate, so recently, some Confy developers are leaving the company, and these complex situations seem to be continuing. Unlike Midjeon, which is currently thriving, it's a bit disappointing that various features of these tools are continuously enhanced, improving the picture quality and convenience. I also have affection for SD, so I feel quite disappointed. If it's not for commercial purposes, it's okay to use it like me, but it's a bit uneasy and scary at the moment. Therefore, it's better to try using Confy UI once to see how it goes. Whether it's feasible or not, there are rumors about it, but I'm a bit hesitant due to capacity issues. So, currently, SD 3 is getting more complicated, and there are many negative reactions in various communities.

We are currently on air. The current situation is like this. It was expected to be bright because it is coming out as open source, but when actually entered, it's a bit chaotic. I think this image is showing this situation well. I'm not sure what will be in the future, but I will also deliver news related to SD. Now, I will wrap up briefly here. Regarding the things about SD3, we will leave more to see. Thank you for watching today's video. Next time, the video will be more advanced, and we will deal with the content here. Thank you.

SD3 Medium이 반갑지 않은 이유

AI 툴이나 관련 정보 지식 팁들을  오프너처럼 열어드리겠습니다.
제휴,제작,강의문의 받습니다. 
문의 메일 : welcronjinjin@gmail.com



SD3 Medium이 반갑지 않은 이유

SD3 Medium이 반갑지 않은 이유

Scripts