Claude Code just made ComfyUI 10x more powerful transcript

Right now there are two different types of ComfyUI users. The ones who spend hours tweaking nodes, looking for the right models, and debugging their workflows. And then there are the people who just tell Cloud Code what they want and let it handle absolutely everything, which is exactly what I'm going to show you in this video. How to make your ComfyUI setup be powered by Cloud Code. And you might be thinking that ComfyUI is too complicated for me.

Well, this is exactly the type of video you should watch because Cloud Code literally builds the entire workflow for you. I'm not over exaggerating. And all you need to do is just describe what you want in plain English. And so you do not need to understand any nodes, any wiring, or any of this technical stuff for you to actually understand how to use Cloud Code to build the workflows for you. And you may be wondering, will these AI generated workflows even work?

Because if you've tried something like ChatGPT, that might have not worked because that's a generic chatbot just kind of guessing nodes. However, we're going to use this Cloud skill, and don't worry, you don't need to understand any of this. So we're basically going to use this skill for Cloud Code to understand how to build ComfyUI workflows so that it's actually good at creating them. And if you have used ComfyUI in the past, you might know that they do have templates and they do have workflows.

However, the problem with using those templates and those workflows is that they're pretty generic and they're just pre-built workflows. So even if you wanted to tailor something to your own setup, you still need to understand how to tweak stuff. However, with this Cloud Code setup, all you need to do when you want to tweak something is speak in English. That's literally all you need to do. You need to tell it what you want to change and it will figure out the rest for you.

And lastly, just before I show you exactly how to actually set all this up, even if you have a terrible PC, all of this is still going to work for you. Because once I show you how to actually set this stuff up, which will only take about 5 minutes, we are going to import the workflows into Comfy Cloud, meaning you can run it on computers that are in the cloud, so they're not actually running on your own hardware. So with all of that out of the way, let's jump into the video now and let me show you how to connect Cloud Code to Comfy Cloud so that I can build your entire workflows for you.

Now of course you're going to actually need to have Cloud Code installed, so make sure you have it installed. And also make sure you have Node.js installed. Both links are going to be down below. So, then just open up the terminal or the command prompt if you're on Windows, and then CD into whatever folder you want to work out of. Then you can just open up Cloud. So, I'm going to do Cloud with dangerously skip permissions.

Then we're going to start the ComfyUI server. Again, I'll show you how to do this on Cloud in just a moment. Now again, you're going to need to have ComfyUI actually downloaded for this to work locally. So, once again, link down below. And then I'm just going to send off this prompt. Now the prompt I sent off is on screen right now. You can copy it. Or like I just said a few times, the link will be below for a base gallery you need in this video.

And so, what we're doing here is we're basically pasting in this skill. And by the way, a skill file is basically this. You make a request, Cloud reads skill.md and picks the right skill for the task, and then inside of one of these skills you have best practices, step-by-step guides, code patterns, tools and library choices, and common pitfalls to avoid. So, that whatever output you're trying to make for whatever use case you're trying to do, it's going to use this skill to create that asset for you.

And that's what basically all of this code and all these settings are that you're seeing on screen right now. So, after I sent off this prompt, it went through a bunch of different things. And so, we have two blockers in the way right now. Number one, ComfyUI isn't running, so I'll ask about that in a second. And number two, Flux 2, which is the AI model that I actually want to use to generate these images, requires a text encoder, which I apparently don't have.

Okay, so given these two blockers right now, I basically said ComfyUI might be in another localhost, so just find out where it is. See, I'm not even telling it what to do. I'm just telling it to solve the problem for me. And then, let's go with option C using when image. I do prefer Flux, but I just don't want to be downloading 5 GB right now. So, I'll revert to another option. And I'm going to send it off. Well, we can see here that I sent off the prompt, and then it did find the port 8000 because it was at a different port.

And now it has suggested creating this long skill.md. So, yeah, we're going to press two and just allow it to create it. And as you can see, there has been some adaptations made. So, obviously changed the port, changed the path, it replaced some of the image models. And now it's saying, "Restart Cloud Code to load the new skill, and then invoke via {slash} ComfyUI underscore local or just mention ComfyUI. So, we're going to do just that by closing the terminal.

So, as you can see, I made a new terminal, and if I press {slash} ComfyUI, it automatically comes up. So, I'm going to click enter, and we're in. Just like that, we have connected Cloud Code to ComfyUI locally. So, now it's asking, "What would you like me to generate? Provide the prompt for an image or a video." So, I'm going to send off this image saying, "Generate an image of tractors blocking a highway due to a protest, police, blue lights, traffic." I'm going to send that off.

Obviously, this is in reference to what's going on in Ireland right now. If you know, you know. Shoutout to all the Irishmen. Okay, as you can see, there is something running here. My PC did slow down for a second. Let's click on view all jobs. Okay, they're running on what's happening, so let's open this back up. And by the way, I'm using Opus 4.6. If you want to check, you just press {slash} model, you can see right now currently Opus 4.6.

Okay, and as you can see, we have our image over here. It took about 483 seconds or 8 minutes. And once it's done, all you got to do is just drag this in or double click it, and it will showcase the entire workflow. Now, this might look cool, but this is just beginning, and this is like the simplest use case I'm going to show you throughout this whole video. So, don't be too impressed just yet, because keep in mind, everything you see here was generated by AI.

Even the connections from each node, everything. It's Yeah, it's kind of mind-blowing. But, I loaded the Diffusion model, the CLIP, the VAE, the latent image, the positive and negative prompts, model sampling, loader model K sampler, VAE decoder, and the save image node. Connected everything up, and I noticed here it also enhanced the prompt that I gave it. So, this is the prompt I gave it. I just said, "Generate an image of tractors blocking a highway due to a protest." And then I gave it a few extra words like police, blue lights, traffic, and it created this prompt.

"Photorealistic scene of a group of large green and farm tractors blocking a multi-lane highway." Like, it basically just went into much more detail, which is why it created this image. So, yeah, this is the image it generated. Obviously, this is nothing special. I mean, this is just ComfyUI. This isn't up to ComfyUI, but I think you get the point that this isn't the interesting part just yet. And look, don't get me wrong, local models and local workflows have their time and place, but if we want the most powerful client generations, we are going to need to use the cloud for faster generations and to be able to use the best open-source models possible.

And that's exactly what we're actually going to dive into now is how to use this on Comfy Cloud and not just in your local workflow. And just before we do that, if you are enjoying this content and if you are learning something, make sure to subscribe down below. It does help out the channel much more than you actually think and it's just a completely free way to actually support the channel. If you do subscribe, you're also going to get much more videos just like this one recommended to you in your for you page instead of some brain rot content.

So, if you did subscribe, thank you very much and let's get back into the video. So, as always, link down below, launch Comfy Cloud this time, and then let's go to platform.comfy.org so that we can actually create an API key. So, up here, click on API key. I'm going to name it subscribe just because I feel like some of you still didn't subscribe. And then we're going to copy this key. Now, I'm going to paste the key in the cloud code.

Normally, you probably shouldn't do this, but again, I'm just showing you like the laziest way possible. And once again, I'm not going to do any technical settings. I'm just saying, "Here is an API key for Comfy Cloud. I want to use Cloud Code to control Comfy UI in the cloud, not locally. Change everything that needs to be changed in order to set this up. And just let me know when it's done so I can reset my Comfy UI Cloud." That's all I'm saying.

Nothing technical. I'm just speaking to it as if I was just telling someone what I want to basically do. And so, just before I create a workflow, I'm going to send off one more prompt, which is this one right here. And basically, what this is going to do is it's going to automatically scan my Comfy UI Cloud so that anything that gets generated will automatically get downloaded to my PC into a certain folder, which I asked for the folder location.

We can obviously give it a folder location if you want to specify where you want the images to go. Okay, so everything looks good. So, I'm going to send off my first prompt. And I basically wanted it to build a workflow with one prompt input and four parallel image generation pipelines, meaning four completely different AI models. I didn't specify which ones because I don't want to think of that. I just want to give it the prompt and I want it to come up with everything else just for for Obviously, if you want the best results, we're going have to start specifying stuff soon, but again, we're working up in the difficulty over here.

So, first I want to just see if it can even do this, cuz if not, then the rest of the video will be useless. Okay, we can see here that it's still running, but it presumably ran some tests because I already have four different images over here. I'm not entirely sure. Okay, never mind, yeah, I am sure because it did say that it wanted to validate everything with a small test to confirm that the models actually load and run on Comfy.

We can also see the four models that it's trying to use, so Stable Diffusion, Flows and Quinn. So, if this actually works, then I am going to send it a very interesting prompt because the next prompt is going to be a lot more challenging and a lot more difficult, so I'm curious to see what it's going to actually output. Okay, and as we can see, we have the images saved, confirmed it, and I already showed you the four different images here.

So, that's pretty cool. But we still don't see the workflow. So, I'm going to ask it where is the workflow? Is it going to be in Comfy Cloud or is it just going to send me the JSON? And okay, so it's saving these as JSON files. And okay, it's saying that it's not going to work in Comfy Cloud, so it needs to basically convert this. So, yeah, I just convert it. Okay, so I didn't know where it was, so I basically just told it open the folder and it did just that.

And yeah, as we can see, we have we have four different ones. Let me just see the most recent one cuz I'm guessing that's the one that works, and let me just drag and drop it into Comfy Cloud. And okay, so this is pretty advanced. I mean, I thought it was going to be a lot simpler just for four images. So, the next workflow I'm about to show you is going to be ridiculously complicated. Maybe it's over complicated, and then I do not know, but let's test it out.

So, I'm going to add a prompt of some sort of ancient Greek philosophy journaling, and just going to click on run and see if it works. Okay, so yeah, one of them already generated. Oh, never mind, it's four, so see more outputs. Okay, so yeah, all four generated. So, that was pretty fast, and I guess this workflow works. And that's the difference between Comfy Cloud and Comfy Local. This generated extremely quickly, and it's created four different completely different pictures with four completely different models.

So, now I'm excited because now I'm actually going to show show like the most interesting use case, and the one which I've been waiting the entire video for. But first I wanted to check if it actually did install automatically and yes, every image that I actually create in Comfy Cloud automatically gets downloaded to my PC, which is pretty cool. I mean, even if you don't end up using this actual use case to create your own workflows, which I don't know why you wouldn't, even that one little script that we just installed, it's very helpful and handy because I mean, you're just saving yourself a little bit of time every now and then by not needing to download these and automatically downloading them to a specified folder and again, you can specify whatever you actually want to download by simply telling Cloud Code which folder you want.

And so this is the prompt. I'm going to basically send it off first so that I can start generating cuz I imagine it's going to take a little bit. But this one's a lot more complicated. Right now, I'm asking for a workflow that takes one text prompt and two input images, a character and a location, and then it's going to generate a four-part short film. Each part will have a different camera angle and visual style. And then I do a quick explanation of how it should work.

So it takes the prompt and it expands the prompt into the four-scene storyline. It's going to use the character image and location as references for consistency. It's going to generate a styled image using the scene description and reference images and then it's going to convert four of those images into short videos using models like 12.2 or Kling and then at the end it's going to combine all four clips into one final video output.

So this is a use case that is actually pretty complicated. This would take quite a while to set up in Comfy Cloud and yeah, it is saying that it's going to require 60 plus nodes and this is good because it basically went into plan mode before completely jumping into this. Cloud locally doesn't expose a hosted chat LM node. If none works, it'll fall back to calling the Cloud API from a helper outside of the Comfy UI graph.

I mean, that's fine. Cleaner model, okay, that's cool. And then it's just giving me more details about the reference images saying that they might not be perfect pretty much, which I mean, it's fine. That's up to the model. Then it's talking more about the styles and more about the video. I mean, all this looks good. Again, it's good that it asked me, but I kind of don't care. I'm just going to say, "Yeah, go ahead." I just want it to do everything for me.

And obviously, this is going to take a bit longer. But again, while this is generating, the beauty of this is that we don't need to actually build a workflow. Even if it does take the same amount of time as you building it yourself, you have that time back, so you can do basically whatever you want. Read a book, watch more of my YouTube videos, or whatever else you want to do. By the way, I've been posting a lot more on Twitter lately, so if you haven't already, make sure to follow me on Twitter.

Twitter is by far the fastest way to stay up-to-date with all these AI news and everything that's going on. So, link is in below. Okay, for some reason, that was way faster than expected. It only took 2 minutes. So, I'm going to create a new workflow. And for some reason, it's not working. I mean, it didn't generate the JSON, so yeah, I'm going to specify that. And I'll see a JSON. And okay, basically, I'm not seeing any nodes.

So, it's a bit concerning. So, I'm just going to tell Cloud Code cuz obviously, I'm not going to be debugging this myself. So, I just told it, "I don't see any way when I load the workflow." Okay, I think it's the same issue that was earlier, which is it was creating the JSON for the local ComfyUI, but not the cloud version. So, it's now going to generate a graph for my version. I have hope. I think that was the issue.

Okay, so yeah, apparently, it converted it. So, let's test this out right now. Okay, so whoa. Yeah, so as I said, this was going to be a much bigger challenge. I mean, you can see the amount of nodes here. It's almost lagging my UI. But yeah, this looks pretty cool. The question is whether or not this actually works now. So, I mean, where do I even start? Okay, I guess that's the issue because I have all these nodes, but do I have to input the prompt into every single node?

Uh I'm not sure. So, let me just take a screenshot of at least this section, open up the terminal, drag the first screenshot in, second screenshot, and third. I'm not exactly sure where to input the prompt. Do I need to input it into every single input field where it says prompt? Yes, that's the annoying answer. Okay, well, whatever. We'll just I mean, we can do a bit of manual work, right? Okay, well, Cloud Code also says that this workflow was designed for the runner.

I'm not sure what this is, but one command, one prompt, and it's done. So, I'm just going to ask, "Can you explain the runner to me in first principles?" Answer in short. So, the runner is a four-step Python script. Upload two reference images to Cloudcode, get back server-side file names. Okay, so basically, is this going to run all through Cloudcode? AKA, I can send you the prompt and you the images, and you're going to run this runner Python script.

Yes, exactly. You just tell me the prompt plus path, and it's going to run it. So, I think that's the better option, to be honest, because as you can see, we do have the workflow. We can just play around with this, but I don't really want to be filling out all of these prompt input boxes, and yeah, this could going to take a lot of effort. So, I think Cloudcode might have just given us a way better solution here. Okay, so my prompt is going to be about a UFC champion.

The four scenes should be him training, eating clean, fighting, and ultimately having mercy to his opponent during fight night. And I attached the two reference images. So, I'm going to send that off. So, I decided to go with Bruce Lee, and I gave him an image of this UFC arena over here. So, we'll see how accurate this is. Obviously, these aren't the best images. Uh the screenshots you attached are only Claude's UI.

They're not the actual files you can read on your Mac. Please save them to a folder, drag both images out of finder, save them into desktop, and send me the file names. Okay, so I just basically saved it there in this folder. I drag and dropped my folder into Cloudcode, and I gave him the names. So, let's paste that, see if he has access. But as I said, these images aren't the highest of quality, but that's not the goal here.

The goal here is to see if this will actually even get this right. Because I don't want to spend too much of my time optimizing the images, making sure they're all high quality, getting all the prompts right, getting all the camera angles. Because first, let's just test it. If it works, then I have this workflow ready. I have everything in place, and all I need to do is just send the prompt, and that's it. Okay, now you can see here that it shows the different styles that it's going to use for the four different clips.

So, for the training one, it's going to use cinematic Kodak. Uh and let me just generate these inside of Midjourney, so you can see. So, I'm just going to do robot in, and then I'm going to send off basically each one of these styles. So, the first style is going to be cinematic Kodak, which looks kind of like this. The second style is going to be noir chiaroscuro, which is kind of like this. Third one is going to be Ghibli anime.

I'm pretty sure you've all seen Ghibli pictures before. And the fourth one is documentary realism, something kind of like this. Okay, it's saying that the status is executing on a cloud. All four scenes are expected to take 15 to 30 minutes. I think that's overkill, to be honest, but the last time it took way longer, even though the images weren't generated. So, I mean, we'll see. I took note of the time, so we'll see how long it actually takes.

I mean, I can already see something is generating here, eight different assets. I don't know why it's saying it's it's going to take so long because I I think it's all generated. And maybe not all, maybe it still needs to stitch the clips together, but we can see here we have a bunch of different assets, which is kind of cool. Let me open my folder and see Yes, we do have a few different things. We have the PNG 1 2 3.

So, yeah, okay, we don't have everything generated yet, but we do have already a couple of videos. Let's open up one of these PNG images and let's see. I mean, even these images are pretty good. I'm not sure what model this is using cuz I kind of just let Cloud Code not even work through the UI, but if I go back here, I can see what models it's using. Oh, yeah, I remember. So, Flux 1 1 2.2 and I think there was like another Flux model in there somewhere.

Yeah, whatever. But just because I already see a few of these, I'm going to say give me an update. Because again, we can talk to Cloud Code about anything, basically. So, if you're not sure about literally anything throughout this video, just literally talk to Cloud Code. Okay, so yeah, this is why you have to ask for updates. Don't just take it by its word saying, "Oh, it's going to take half an hour." cuz it took literally 2 minutes and it said that the job succeeded, but that we only generated three MP3s and three MP4s because scene four was missed.

So, now it's going to recover the missing scene four and regenerate it. So, yeah, regenerate scene four and finish the final output. All videos stitched together. Okay, so apparently stitched all the videos together. So yeah, this is the final video. Let me just put it here and play it. Yeah, so there's no audio. Okay, the character consistency, I mean, first of all Okay, let me just play it so you can actually see what's going on first.

Okay, there's a few things that are wrong here. Kind of. I mean, this part, I'm guessing, is the part where I said, "Have mercy on the opponent during the battle." This part looks like training. Oh no, this part looks like training. And here we can see he's falling down or losing the fight. I kind of forgot what my prompt was, to be honest. So yeah, it should be training, clean meal, fight, and then mercy. So it did I mean, I mean, look, this is not entirely the fault of ComfyUI.

This is kind of the fault of the actual models themselves, I presume. Maybe the prompts were written wrong. Maybe Cloud Code didn't actually understand how to write the prompts, I don't know. But I presume this is just a fault with the models because we do have the entire workflow here. I'm not going to fact-check to see if this is actually all correct and if everything is in the right place or not cuz it's going to take too much time.

And the point of this video was just to show you that you can use Cloud Code to actually create these workflows. This one might be a bit overkill right now, but again, I'm sure if I spent a little bit more time just explaining that this is overkill, let's simplify this, remove any nodes that are unnecessary, or explain the process to me so that I can understand it even better, then I would do just that. And if you did enjoy this video, then I'm sure you would enjoy a video about Cloud Code and re-motion, which is currently the best way to edit videos with AI.

So if you want to see that video, just click somewhere up here and I'll see you there.