Transcription
Higsfield is now integrated with Claude, which means that the days of manual AI prompting and switching between tabs are over. But it's not all good. I made some pretty expensive mistakes when testing it out, and I'm about to share those with you so that you don't waste your credit.
Hakesfield sponsoring this video, and their new Claude MCP connection can ideulate storyboards, create cinematic videos, test image models, and tons more. All using the intelligence and speed of Claude to create the best results in less time.
So, the first question is setup. How do you get this thing to work? First, go to Higsfield and then navigate up here to the MCP and CLI tab. On this page, make sure that Claude is selected and copy this link right here.
Now, you're going to want to open up Claude or sign up if you don't have an account yet. Then, move up here to the sidebar menu. Navigate down to customize and then click connectors. Here, click the plus button and then click add custom connector. Now, we'll name this connector Higsfield. And down here is where we paste the link that we copied earlier. This will open up an authentication flow to link your accounts. We'll click through this. And just like that, Higsfield is now linked to Claude.
The page we're on now, though, is really important. These are all of the permissions that decide what things Claude can do autonomously with Higsfield and what things you want to require it to ask permission on before acting. All these things like showing generations and exploring models, I recommend setting to always approve just to keep things smooth. But I highly recommend requiring Claude to ask approval before generating [music] images and videos.
To show you why I say that, and also one of my favorite ways to use this agent, let's test it out on creating a whole AI cinematic sequence from scratch. I'm just going to tell Claude that I want a one minute long action film with [music] an anime space army fighting aliens on the moon. And just like that, the agent has given me a whole storyboard with different shots to make up the film. I think this whole thing looks good, but obviously it hasn't given us real storyboard images yet. So, I'll ask it to use Higsfield to mock up those references. Immediately, it's identified that Cinema Studio 2.5 might be a good model to create this style of image. And I okayed the generation, which resulted in this [music] result. Here's the next image result that it gave me. Again, spot-on.
Now, this is what the generation flow looks like. You can see it's got all of the parameters in place for the image model. And this right down here is the prompt that it's going to use. Now, before it uses my Higsville credits to generate it, I have to click allow right down here. It adds another manual step to the process, but this is a great safeguard to make [music] sure that you're not using way more credits than you intend to use. The third image that it just gave us also looks good to me. So, I think I'm just going to click always allow and let Claw do the work on all of the rest of the images. We'll see how that works out.
Now, here's what I'm talking about and why you want to be careful. The agent totally lost the style of the film in those next generations. This isn't exactly Claude's fault. I was just kind of letting it do a lot on its own without very specific instructions. So, I'm going to be more intentional about this and say, "Hey, we need to regenerate these." And instead of using Cinema Studio, I want you to use Nano Banana Pro and use image 2 as a visual style reference for the rest of the pictures to keep that style consistency. With that, it regenerated all of the images and all of these look way better than what we got before. So, while the agent can help automate a lot of the steps in the process, you still need to know what you're looking for and the strengths [music] and weaknesses of other models. It's going to give you a much better result that way.
So now these pictures look cool and I'll just ask it to use the original film idea as well as the images to create videos with Sea Dance 2.0. Unfortunately, I made another mistake here and you're about to see it. Claude gave me results for the first three shots in the sequence, but I'm really not very happy with them. They feel lifeless because Seance does much better when given very descriptive prompts and then enabled to make multi-shot videos instead of single shots based on individual start frames. [music] Knowing this, I just asked Claude to plan new prompts for the scenes and combine them together into four different generations for a [music] better result. The AI totally understood what I was asking for and it gave me these prompts, which [music] I wanted to make more descriptive. So, I just had it expand them a bit.
Now that I'm happy with these prompts, let's have it generate a whole batch of sea dance videos. But, we'll start with just the first one to make sure that everything goes according to plan. A minute or two later, we've got our result. >> Dead silence of deep space. Aside from the deep space narration, this actually looks really good. So, uh, let's just have the rest of the shots go through. Now, Claude has tried all three generations, and the last two came through, okay, but the first one was blocked for some reason. So, let's just try it again.
Okay, now we have a result here. I know it doesn't look like a result. We've got this sensitive content warning on it, but this is just a temporary bug and I'll show you how to fix it. When you get a warning like this, just open up Higsfield's website and navigate to your video generation tab. Here you'll see all of the generations that we just did with Claude. But importantly, you'll also see this one right here that says rights verification required. This is the video that wouldn't show up through Claude. If we just click through this and confirm rights, the video will become visible here on Higsfield. And if you just go back to Claude and reload the conversation, you'll see that the video is now visible here inside Claude as well.
Now, here's our finished sequence. >> Dead silence of deep space. form defensive perimeter. It's not the best AI video sequence I've made, but at least it serves as a good example of how this workflow works through Claude.
Powering Claude with Higsfield goes way beyond this. Just imagine anything that you can utilize a powerful LLM like Claude for and then also the power of Higsfield. You can generate images from any AI model. I can even chat with the agent and have it create product mock-ups and stuff with GPT image or Nano Banana, even though I'm not using Chat GPT or Gemini. Check this out. I want to brainstorm a men's shampoo product and brand. Claude, give me three product name and brand ideas for a hair shampoo for men. Here we go. We've got Velt, Ridgewood, and Brink. Not bad. Now, create a optimized GPT image 2 prompt for each one of these to make a mockup of the product bottle. Three optimized prompts just like that.
Now, I could manually copy and paste these things into Higsfield and get generations, but that seems like a lot of effort. So, Claude, use Higsfield to mock up each bottle using these prompts in their own GPT image to generations and [music] make the generations 3x4. Here's Velt. Here's Ridgewood. And here's Brink. I like Ridgewood best, but it looks kind of plain right now. So, let's add some like nature inspired elements. Cool. Add Bigfoot. That's what I'm talking about.
Finally, let's talk downsides. Because whenever something awesome and new comes out, there's a whole slew of downsides that comes with it. And that's why you guys watch me because I'm going to show you the truth and tell you about the things that aren't so great. The most glaring downside here is something that I mentioned before, but it's credit consumption. Credit consumption is basically impossible to keep an eye on when using Higsfield through Claude. Now, there are things you can do to mitigate the damage. Like you can ask Claude to only generate in 720p to save credits or to do 2K generations instead of 4K images by default. And of course, you can set it to ask permission before doing any actual generations. But there's still no indicator of how many credits each generation is going to cost or how many credits you have left on your account when you're using Claude. Both of which you'll find if you're just generating on Higsfield itself.
The only other real downside that I see to using Higsfield through Claude is that Higsfield has a really helpful interface with features like image linking, sole character selection, and even aspect ratio selection, all of which can be controlled through Claude. But you need to already know that these choices are there if you're going to instruct an AI agent to use them. So for beginners, I highly recommend just using Higsfield and getting experienced with that UI and all of the tools and features that they have accessible and then converting over into using it through Claude just so you have a better idea of how things work. This will also give you a lot more valuable insight into credit costs and know how to manage those just kind of inherently when you're using Higsfield through Claude.
That said, this new feature with Higsfield is so incredibly powerful and really the things that I showed are just scratching the surface of all the stuff that you can do. And I'm really excited to see how you guys are able to use it. Let me know if you have any questions. And if you're interested in seeing my in-depth tests on Seance 2.0, know, which is by far the best video model that we have to date.