The Gemini + Higgsfield System for Consistent AI Videos

0:00There's one real reason your AI videos keep morphing and looking weird, and it's that the description of the perfect video only exists in your head. But the moment you build a system that holds that description outside your head, the morphing stops completely. And that system is a custom Gemini gem used with Higgsfield. And in this video, I'll show you exactly how it works. So every AI video you make comes out consistent from the first shot to the last. The reason AI video keeps coming out different every time is that the tool has no memory of your last shot. Every new generation rebuilds the face, the room,

0:29and the light from scratch, working only from the words you give it in that exact moment. So you can paste the same description twice and still get two different people because the model is filling the gaps both times and filling them differently. The usual response is to keep changing the prompt, piling on more detail, trying to pin the character down word by word. But that only ever works for a single shot, and the second you move to the next one, you're describing it again from memory, and

0:51small differences slip in every time. The real fix is to stop holding that description in your head and hand it to something that writes it out the same way on every shot. That something is a custom Gemini gem, and it controls the entire look of every video you make. It writes one locked description that covers the character, every location, and the overall style, and then it writes every Higgsfield prompt for you from that same description, which is usually where consistency quietly falls apart. I've put this exact gem in the description, so you can grab it and follow along with the same one. That's the system, and the whole thing starts

1:21the moment you hand the gem your project. You open a chat with the gem and describe the video you want in plain language, and a few sentences is enough. For this one, I'm making a short video of myself making a morning pour-over coffee and then relaxing with it out on my balcony with the same person in the same rooms the whole way through. I give it the basics, around 40 seconds, three videos, 16 by 9, a clean photorealistic look, the two locations, and the fact that the character is still me with my face uploaded in Higgsfield. And then it confirms the defaults it's working from,

1:50three videos at 1080p with the audio on. Then in a single reply, it hands back the entire plan for the video. The first thing it writes is the locked description itself, the one style line, the full character description for me tagged as @parker, and a separate description for each location tagged at kitchen and at balcony. Those tags matter more than they look because they're the exact handles you reuse in every prompt later, and reusing the same handle is part of what holds the look together. Next, it writes a ready-to-paste image prompt for each of those three references. So, you have the precise text to generate me, the

2:20kitchen, and the balcony as images. After that, it lays out a shot map, which is a plain breakdown of what actually happens in each of the three videos, so you can read the whole story before you generate anything. And then it writes a complete video prompt for every video split into individual shots with the references already tagged inside each one. So, from a few sentences of description, you get the locked look, the image prompt for every reference, the shot plan, and the full

2:44generation prompt for all three videos. Every piece of that is written from the same description, which means the consistency is built in before you've generated a single frame. There's nothing left for you to redescribe later, and that is the entire reason this holds together. Everything you make from here pulls straight from this one reply, and if you want to change any part of that plan, you never touch the prompt yourself. You tell the gem what you want changed in plain language, and it rewrites only that piece while everything else stays exactly where it was. For the first video, I decide I want it faster with more close-ups and

3:14the actual sounds of the coffee being made. So, I tell it to make the first video fast-paced with tighter shots and the ASMR sound of the grinding and the pouring. It rewrites just that one video prompt, and now the shots are closer in, and the grinding, the filter, and the pour each carry their own sound. What it does not touch is everything else. So, me, the kitchen, the style, and the other two videos all stay locked to the same description they had before. That's the part that keeps you in full control without any risk because you can redirect the creative as much as you want, and the project never slips out of

3:42sync behind you. The same thing works on the image prompts. So, if a reference needs a change, the gem rewrites that prompt, too, which becomes useful the second you start generating the actual images. Now, you take the image prompts the gem wrote and turn them into the actual reference images inside Higgs Field. I start with myself. In Higgs Field, I open the image generator, choose GPT image 2, and upload my real face as the reference image. Then I paste in the character prompt the gem wrote, which is a full-body shot of myself in a plain neutral pose against a clean light gray background with the face pulled straight from the upload. It

4:13hands back a clean character sheet of myself, and I save that as an element, which means every video from here can load the exact same image instead of generating a new one. Then I do the two locations the same way, still on GPT image 2 with no face needed this time. I paste the kitchen prompt for an empty modern kitchen with the kettle, the grinder, and the pour-over setup, then the balcony prompt for an empty balcony with a chair and a railing looking out over a city. I save each of those as its own element, too. So, now there are three saved references: myself, the kitchen, and the balcony. And these are

4:43what every generation will load. Saving each one as an element is what lets every video pull the same person in the same rooms instead of the tool guessing at them again on each generation. If one of them comes back slightly off, you don't fix it by hand, you send it back to the gem. When I decide I want me in a yellow T-shirt instead, I just tell the gem and it rewrites the character image prompt with that one change and leaves

5:03everything else exactly as it was. Adjusting after you see the real images is normal on any production, and because the gem still controls the description, that fix carries through everywhere the reference gets used. All of the image and video generation in this system runs inside Higgs It's the one tool here that turns your saved elements into both the reference images and the finished videos. The link to it is in the

5:23description if you want to follow along. Now, the saved references turn into actual video, and you do that in Higgs Field with Seedance 2.0. The first video is me in the kitchen grinding the beans and setting up the pour-over. I load me in the kitchen as references, set it to 15 seconds at 1080p 16 by 9 with the audio on, and paste in the fast-paced prompt the gem wrote. Because both the character and the room are loaded from saved references, I match my character sheet to the kitchen rather than having

5:47the prompt rebuild them from scratch.

6:05The second video is the part that makes this whole system work. It's meant to pick up exactly where the first one ends, so you generate the first video before you touch the second. Then you open upload media, go to your video generations, select that finished first video, and save it as a reference. You load it into the second video alongside myself in the kitchen, and the prompt itself opens with the line continuous shot from it video one. That line tells SeaDance to carry the motion straight across the cut, so the two videos join with no visible seam, no jump in his position, the light, or the objects on

6:33the counter. I keep the same settings, 15 seconds at 1080p, 16 by 9, audio on, and paste the second prompt where I pour the water over the grounds, watch the coffee bloom, and lift the finished mug.

6:58The third video moves him out to the balcony. So this time I load myself in the balcony, drop it to 10 seconds, and paste the last prompt where he steps outside, reacts to something past the railing, pulls the chair around, and sits down with his coffee.

7:17All three videos were generated separately, but both rooms and I still match from the first to the last, which leaves only one thing to do. The last step is assembly, and for that I use CapCut. I drop the three videos in order, the kitchen grind, then the brew, then the balcony, [music] and because the first two were made as one continuous shot, they sit together

7:33without a noticeable cut between them. Then I add a background track underneath and balance it so the coffee sounds still come through over the music. [yawns]

8:18And that's the finished video. Around 40 seconds of one consistent person in two consistent locations built entirely out of separate generations. Nothing about me or either space changes from the first second to the last. That's the whole reason for running everything through the gym in the first place instead of prompting each shot on its own and hoping it matches the one before it. Once every prompt you generate comes from one locked description, consistency stops being something you force shot by shot and becomes the default for the entire project. And the reason every shot in that video matched is the saved references inside Higgs Field. The same

8:49me and the same two rooms loaded into every generation. That one habit, reusing your saved elements instead of describing them again, is what keeps an entire project locked together. So, when you're ready to build your own consistent video, the link to Higgs Field is in the description and you can have your first references saved in the next few minutes. Thanks for watching

9:07and I'll see you in the next one.