Hello, I'm Sanai of SanAI Today, through text Video generation, updated to 5B model. Let me explain about COGVIDEO. First, run the Stability Matrix. If you don't have the Stability Matrix and run the ComplyUI. Please update the latest version of the ComplyUI. Let's start Comedy UI. Go to the Manager's Install custom nodes. If you search Cog here, Kijai has the CompyUI-CogVideoXWrapper. Please install this. If this custom node has been installed, please update all nodes through UPDATE ALL. Here's the CompyUI-CogVideoXWrapper. Let's take a look at what COGVIDEO is. If you come in here, it's connected to Hugging Face. It's COGVIDEOX-5B. When you get these prompts, these videos are created. The video we're going to watch today is a video of a dog running. I'm going to make it into text. And the baby who's sleeping can be seen through the prompt.

you can see that it was created. If you look here, it's a single GPU VRAM. If it's BF16, you need 26GB. I'm using 12GB VRAM. If you set it up, the quality would be a bit poor. I'll explain how it's possible to create it. First of all, if you installed the custom node. You have to execute the requirements here. First, in the Data folder, in the Packages, in the ComplyUI. If you haven't used Matrix. Please come to the ComplyUI folder. Here are custom nodes. The CogVideoXWrapper installation is complete in this ComplyUI. If you have restarted the ComplyUI. Like this, we have the CompyUI-CogVideoXWrapper. There are requirements here. If you type cmd here. The command window opens like this. You've done it a lot, right? Please run pip install -r requirements.txt. Then, the installation is complete. If you look here, in the samples. There are three Json files. There are three workflows.

These two are the workflows used in the 2B model. Let's take a look at the 5B model. First of all, in CompyUI, I'll drag this JSON file and put it in. Then, Workflow brings it in and loses. I'm going to bring in CLIP here. CogVideo TextEncode를 통해. This CogVideo sampler creates an image. And through this CogVideo Decode, I'm creating a video. To use this Video Combine Video HelperSuite requires this custom node to be installed. I think everyone installed this. Here's the ComplyUI Video HelperSuite custom node. Please install this. If it's not installed, please install it. Please update your custom node through UPDATE ALL. In this state, in CLIP name Google's T5. There's this version of FP8 text encoder. If you don't have this. Install Models exist in this Manager. If you search for T5, you can find many things. You can install Google's T5 here. Before, SD3 or. Or if you've used the FLUX model.

I think this text encoder is already installed. This text encoder was used by SD3 and FLUX. This is the T5xxl_fp8_e4m3fn.safetensors file. If you set this up, I'm going to create this video of a Golden Retriever running. Here are these settings. If the number of these settings is different, please update the custom node. You can choose 2B and 5B models here. I'm going to choose the 5B model. We have FP16, FP32, and BF16. If you set it up like this, I'll use the rest as the default. But in the case of VRAM, as you can see, he said he needs 26GB. But we're going to use the encoding of the FP8 version. Please change this FP8_TRANSFORMER to True. If the graphics card isn't very good, you have to use this CPU. If that happens, the speed and quality of the video will decrease a lot. You'll be able to check it out.

I use about 12GB of VRAM. If you use this much, you can create a video with this version of FP8. And here's the video size: 480 x 720. So far, we can't create a higher resolution video than this. I'll try to use this DPM for this scheduler, too. And you have to set up one more thing. If you have enough VRAM, you can proceed. We don't have enough VRAM. I'm going to set this to Enable VAE Tiling here. And I'll set Enable VAE slicing to True. In the case of 96, 96, 0.083, and 0.083 set here, it's the default part. If you change the value, should I say it's a little out of focus? That's why I think it's the best value. You don't need any other settings. If you need to save a video, you can set it up here. I'm going to create an MP4 format video like this. There is a (Down) load CogVideo model here at first run.

If you do this, the model will be downloaded. About 12 or 13 gigabytes of capacity will be downloaded. You download several files at once. The models are in the ComplyUI folder, in the Models folder. A folder called CogVideo is created. If you run 5B here, in the 5B folder. Various files are created like this. If you look at this, it's downloaded about 10.7GB. And even if you use the 2B model, you download the model like this. It's about 3.5GB now. When I first start, I download all of these models. Please wait for a moment. If this (Down) load CogVideo model does not work. If you look here, you can see that the download is going on. It takes a lot of time. First of all, we're doing this. I'm recording this video right now. This memory overflow may occur. Let's move on. You can see that GPU and VRAM are used up to 100%.

It took about 13 minutes and 3 seconds. A puppy running like this has been created. What do you think? We created a video. But since I used FP8_TRANSFORMER, the video quality is a bit lower than it looks. So if you have enough VRAM, you can create a clean video like this. And one disadvantage is that ENABLE VAE TILE. When I use this, I can't feel it when I see it up close. If you look from far away, there's a line like this. You'll be able to see it a little bit. Let me introduce a way to solve this problem. And one more thing. It seems like it's cutting off a bit, right? To solve this problem, we're going to have to make a video combination. I need to add one image interpolation, right? If you search VFI, there are a lot of Interpolation Nodes. In order to use this ComplyUI FRAME INTERPOLATION. NODES needs to be installed. If you search for Interpolation in Install Customodes.

Here's the custom code for Comedy UI FRAME INTERPOLATION. Please install this and update it later. Please restart the ComplyUI. While interpolating VFI images, I'm going to use RIFE VFI custom NODES. It's as easy as that. I'm going to put in the image that was created here and the image that came out of this place and create a video combination. It says 2 here. An image frame is added between the frame and the frame. There are more images that are created here. If you set it to 8 frames, the video will be slower than this video. So, for example, if you increase it by 3 times, set it to 24 frames or 23 frames. Let's do it again. The image has already been created. It's less choppy than before. I'll try to get rid of the string here. I'll save this video to get rid of this line. I'll group them into groups and disable them. I tied it into groups. I'll do Bypass Group Nodes.

I'm going to bring in a NODES. I should call the video now, right? Load Video NODE. Please add a custom nod for Load Video (Upload) here. I'm going to choose here and put in this video that I just created. That's where the image comes from. I'm going to do Upscale with the image that came out here. If you search Upscale, here's the CR-UPSCALE Image. Custom node called ComplyUI-Comply-Custom must be installed. If you look at this, it's Compyroll Studio. If you search Compyroll, there's Compyroll Studio. If you install this, you can use this custom load. Leave it like this. Here's the Upscale model. Well... It's a video from the live-action background. Real ESRGAIN. I'm going to use Upscale. You can adjust the Rescale_factor to the size you want. When the image comes out here, I'm going to make a video combination again. I'm going to add Video Combine to the CompyUI-Video Helper Suite.

The frame will be the same here, right? This video has 24 frames. So, I'm going to set the video combination to 24 frames. I'm going to use MP4 instead of GIF. And if you need to save it, you can set the Save Output here. You can set the route and name you want here. In my case, to distinguish this video separately, I'll put it in the SanAI\test folder. It's going to be in the test folder. If you create it in this state, I'll take this video, do Upscale, and create it. If you look here, you can see the line well. You see this line, right? This part will be removed and the video will be created. Let's wait for a second. The video became Upscale. If you look here, you can see that the plaid has shrunk a lot. But I think there's still some reflection in the water here. But as before, you can see that the line is not marked. As these videos became Upscales.

You can see that these plaid have disappeared. But when you create an image, if you look at the bottom, you can see that it's a little thick. To improve this part, there's one more way to look at it. Here's VEnhancer. For VEnhancer, this unnatural video, you can show it naturally like a real video. And if you look here, please take a close look at your face. As for the face, it's a bit crushed, right? It's the same for eyebrows, too. If you use this VEnhancer for the crumbled parts, you can act out the video like this. I tried to use this VEnhancer several times. I think Kijai is still in the development stage. And when I tested the VEnhancer with this model, VRAM takes up a lot of space. To support FP8 four days ago, I think that's how I'm developing it. If you improve these areas with this VEnhancer, I'll share it with you. If you come out like this, I think it's going to be a very clean video.

We've looked at Text-to-Video. There was also the next workflow, right? If you enter the COG Video Wrapper at Custom Node Video-to-video workflow. Please put a video here too. And through the text here. It explains how to change it. I've used this text prompt for Jeepcha and create a video. Please test this part yourself. To reduce the volume. You have to use FP8. When I used the 2B model FP16 and FP8_TRANSFORMER. I hope you can use this as well. Through the text and video. When you create a video-to-video. If you check the results. There's a video like this down here. When you create it with these settings. Are these Jeepcha videos these bears? It's modified as it says on the prompt. You can watch the changed video. The video is a bit thick here too, right? The way to improve this part is. If you use VEnhancer. I think I'll be able to create a very soft video. Then through this text-to-video. I'm going to watch some of the videos that I made for testing.

I'll wrap it up now. It's a puppy running like this. And he's begging for food. And a video where you want to go for a walk. And a video of me taking a nap outside during the day. A video of someone playing with a toy. And a video of me eating. This time, it's a cat. A video of a cat resting by the window. A video where you're eating. A video of me sleeping on the sofa. A very cute kitten. It's a video of me playing on the sofa. All of this was written in text. This text prompt is very easy to create. I created this with ChatGPT. There's a sample like this, right? I've used 9 prompts here. These 9 prompts. If you look at COGVIDEOX. There are each of these prompts. There are eight here. If you look at the bottom of the VEnhancer. There is a prompt related to this boat. Please put all of this in. Please let ChatGPT know that there's this prompt. I'm asked to write a prompt in this style.

And the cat's daily routine. I made a request like this. I got these prompts. And the scene where Golden Retriever runs on the beach. That's why I requested it through ChatGPT. This is the video. It's a video of me running by the beach. Please express Golden Retriever's daily life. I asked for a lot of prompts and created them. Please make a video like this. If you like graphic cards. Please turn off this FP8_TRANSFORMER. Underneath here, enable vae tiling. Please turn this off and test it out. Let me tell you one last thing. If you come into the CogVideo Hugging Face. Hugging Face Space. You can test it right away if you come in here. I'm going to add a golden retriever prompt. It increases the resolution. With 8fps increased to 16fps. If you press Generate Video. A video will be created in a little while. You can create a puppy that runs right away. There are a lot on the bottom too. We created a lot of them.

There's a video of me painting like this. There's a video of a kid running out. In this case, it came out kind of weird. I like this one. It's a video of painting on the beach. But if you look here, you can see a lot of lines, right? This is a stage I haven't done. Upscale yet. The video where I paint is well expressed. Also, the movie-like scenes are well expressed. What I liked the most was. The video of me holding a baby like this looks like a real person. That's what we created. Besides this. I expressed a boat floating on the carpet. And like this. I tried to express a child crying on the battlefield. Doesn't this look very natural, too? It feels like an old movie. You can see it moving very naturally. We looked at CogVideoX-5B. As I said before. When this VENHANCER is available. Let me introduce it again. If this video helped you. Subscribe, Like, Set Notifications. And please sign up for membership.

I'll be back with a better video next time. Thank you.

🎬 ComfyUI 로컬 무료 비디오 생성 AI CogVideoX-5B #comfyui #texttovideo #ai

안녕하세요! "San AI 연구소"에 사나이(SanAI)입니다. 여러분과 함께 생성형 AI의 신비롭고 흥미로운 세계를 탐험할 가이드입니다. 이 채널은 생성형 AI 기술에 대한 깊이 있는 정보를 제공하고, 특히 Stable Diffusion 같은 최신 AI 도구들을 활용하는 방법에 대한 가이드를 제공합니다.

여기서는 AI 기술의 기본 원리부터 시작하여, 실생활 문제 해결에 이르기까지 다양한 주제를 다룹니다. Stable Diffusion을 사용하여 놀라운 이미지를 생성하는 방법, AI 모델을 교육시키는 과정, 그리고 이러한 기술이 우리의 일상과 업무에 어떻게 혁신을 가져올 수 있는지에 대해 알아볼 예정입니다.

최신 생성형 AI 동향을 따라잡고, 실용적인 Stable Diffusion 튜토리얼을 찾고 계시다면, 여러분은 올바른 곳에 오셨습니다. "산AI 연구소"는 기술 애호가, 디자이너, 개발자, 그리고 모든 창의적인 사람들을 위한 공간입니다. 우리와 함께 생성형 AI의 무한한 가능성을 탐험해 보세요. 구독하고 알림을 켜서 새로운 발견과 학습 기회를 놓치지 마세요!

Forge를 설치 안하셨다면 : 
https://youtu.be/xgoTQP9Sm7M

ForgeUI & A1111(WebUI)  기본 설정 및 필수 설치 확장 프로그램 : 
https://youtu.be/LC1cu9BEvxA

Stability Matirx 를 모르신다면 :
https://youtu.be/3EBCz0oU3vA

이메일 문의 : music.with2me@gmail.com



AICOMPANY@XMI66V

shortsmusic@wtHjwx


🎬 ComfyUI 로컬 무료 비디오 생성 AI CogVideoX-5B #comfyui #texttovideo #ai

🎬 ComfyUI 로컬 무료 비디오 생성 AI CogVideoX-5B #comfyui #texttovideo #ai

스크립트