Big Idea

The argument in one line.

A video carries two separable layers, who is in it and how things move, and because AI content filters only police the who, stripping identity out while keeping motion intact gets any face-swap past the block.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use AI video tools like Seedance, Higgsfield, or similar generators and have hit an IP or face-recognition block.
You want to understand the actual mechanism behind a face-swap trend video instead of just buying the wrapper app that sells it.
You're comfortable chaining open-source models (depth estimation, segmentation, audio separation) through an AI coding agent.
You want a free, self-hosted alternative to a paid face-swap subscription product.

SKIP IF…

You're looking for a no-code, point-and-click tool — this requires an API key, a coding agent, and some manual setup.
You have no interest in the technical 'why' and just want the finished video.

TL;DR

The full version, fast.

Seedance 2.5 blocks a face-swap when it recognizes a famous face or copyrighted music, because those are the only two things the filter actually checks. The creator strips both: Depth Anything converts the source video into a colorless depth map that keeps every motion but hides every face, while vocal isolation and a pitch shift remove the copyrighted soundtrack without losing the words the AI needs to sync lips. A depth mask alone returns the wrong body and a gray void for a background, so he composites the depth-masked people back over the real footage to fix both. He then automates the whole four-move process, plus Demucs, SAM 3 segmentation, and draft-mode generation, into an 11-step pipeline he gives away as an open-source agent skill, with a face-mesh variant for single-subject talking-head videos.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:38

01 · Intro

Sirio shows a face-swapped clip of himself in a movie scene, then demonstrates the same request getting blocked when tried directly in Seedance 2.5.

00:38 – 03:32

02 · Why Your Video Gets Blocked

He frames the core idea: a video carries a 'who' (face, voice, music) and a 'how' (motion, timing). Seedance's filter only checks the who, so hiding identity while preserving motion is the whole trick.

03:32 – 05:13

03 · Depth Anything Explained

He runs the source video through Depth Anything to generate a colored depth map, stripping faces while keeping every movement intact.

05:13 – 07:07

04 · Isolating the Vocals

Since muting breaks lip-sync, he pitches the audio up and isolates the vocals in CapCut, removing the copyrighted music while keeping the spoken words the AI needs.

07:07 – 08:18

05 · API Key Setup

He generates an Enhancor API key for Seedance 2.5 and has a coding agent store it safely in an .env file rather than pasting it into chat.

08:18 – 09:38

06 · Draft Mode Explained

He explains Seedance's draft mode: cheap, repeatable low-res previews tied to a task ID, with the final high-res render paid for only once.

09:38 – 11:02

07 · First Test

He submits the depth mask and two reference photos for a first generation using the draft workflow.

11:02 – 13:47

08 · Two Fixes

The first result has the wrong body and a gray background. He fixes the gray by describing the room to the agent, then fixes the wrong body by compositing the depth-masked people back over the original footage.

13:47 – 15:04

09 · My 11 Steps Workflow

He lays out the full automated pipeline: Demucs for vocal splitting, Rubberband for pitch shifting, Depth Anything, SAM 3 segmentation, compositing, vocal reinsertion, manual review, then Seedance draft and final generation.

15:04 – 15:47

10 · Get the Workflow for FREE

He announces the open-source project and skill package covering the entire pipeline, free to download and set up with a coding agent.

15:47 – 18:23

11 · Workflow Live Demo

He runs the finished interface end to end: upload, automated prep steps, mandatory visual approval of the mask, character selection, and final 1080p export.

18:23 – 20:37

12 · Face Mesh Workflow

He demos the faster alternative for single, centered subjects: a 478-point face-tracking mesh that hides identity while preserving exact facial motion.

20:37 – 21:46

13 · What to Build Next

He recaps the core principle, keep the how and throw away the who, and suggests extensions like adding a transcriber or using your own voice to build a full personal avatar.

Atomic Insights

Lines worth screenshotting.

Every video carries two separable layers: the 'who' (face, voice, music) and the 'how' (motion, timing, lip movement), and AI content filters only police the who.
Seedance 2.5 blocks a face-swap request for exactly two reasons: a recognizable face or copyrighted music, nothing else.
Converting a video to a colored depth map strips away identity entirely while preserving every frame of motion, which defeats face-recognition filters without losing the performance.
Muting a video to dodge a music-copyright filter breaks lip-sync, because the AI model needs to hear the actual vocal track to animate the mouth correctly.
Running a track through vocal isolation and a pitch shift removes the copyrighted music and the identifiable voice while keeping the spoken words the AI needs for lip-sync.
A depth mask alone reconstructs the wrong body and a gray void for a background, because the generator has no information about what the environment actually looks like.
Compositing the depth-masked people back over the original background stops the model from blindly tracing the depth outline, which fixes the wrong-body problem.
Draft mode renders a disposable low-resolution version tied to a task ID, so cheap iteration should happen there before paying once for the final high-resolution export.
Re-uploading a downloaded draft as a new reference file breaks continuity, because the draft is tied to its original task ID, not the video file itself.
The entire manual workaround collapses into an 11-step pipeline: split vocals, pitch them up, generate a depth video, segment people out, composite them over the background, reinsert the vocals, then generate.
A face-mesh overlay tracking 478 facial points is a faster alternative to full-body depth masking when only one centered, camera-facing subject needs to be hidden.
Two or more people moving around favor the depth-mask method over face mesh, which is built for a single centered subject.
Stacking a face mesh on top of a depth-masked cutout combines both identity-hiding techniques for a more failure-resistant result.
The workflow deliberately gates the expensive generation step behind a manual visual check for mask holes around the hair and hands before any money gets spent.

Takeaway

A video's motion and its identity are two separate things, and only one gets filtered.

WHAT TO LEARN

AI content filters check who is in a video, not how it moves, so isolating and discarding identity while preserving motion is a general-purpose way to get past a block without losing the performance.

02Why Your Video Gets Blocked

A video carries two separable layers, who is in it and how things move, and an AI filter that checks for identity won't catch a video with the identity removed.

03Depth Anything Explained

Depth estimation turns people into featureless, distance-based shapes, stripping faces while keeping every frame of their movement intact.

04Isolating the Vocals

Muting audio to dodge a copyright filter breaks lip-sync because the model needs to hear the actual words; isolating vocals and shifting pitch removes the music while keeping what the model needs.

06Draft Mode Explained

Low-cost draft or preview modes exist specifically for iteration; treat the first several attempts as disposable and only pay for full resolution once satisfied.
A generated draft is tied to its task ID, not the output file, so re-uploading a downloaded draft as a new reference starts over rather than continuing.

08Two Fixes

A depth mask alone loses body shape and environment detail, because the generator has no information beyond silhouette and distance, so the real background has to be reintroduced separately.

09My 11 Steps Workflow

A repeatable multi-step manual process is a strong signal that it should be automated into a single pipeline or agent skill rather than redone by hand each time.

11Workflow Live Demo

Gate any step that costs real money behind a manual human check, since automated generation steps are unpredictable enough to waste spend on a bad result.

12Face Mesh Workflow

When choosing between two workaround techniques, match the technique to the scene: single centered subject favors a lighter-weight method, multiple moving subjects favor the more thorough one.

13What to Build Next

The same identity-versus-motion principle that defeats an IP filter generalizes to other goals, like building a personal avatar driven by your own voice and movement.

Glossary

Terms worth knowing.

Seedance 2.5: An AI video generation model that can swap a person's likeness into a reference video, but blocks a request outright when it recognizes a famous or copyrighted face or soundtrack.
Genjutsu: Higgsfield's feature that swaps a user's photo into a template video without tripping Seedance's built-in face or music filters, built on the same identity-stripping idea taught in this video.
Depth Anything: An open-source model that converts a video into a colored depth map showing how near or far things are from the camera, turning people into featureless silhouettes while keeping their motion.
SAM 3: A segmentation model used to cut people out of every frame of a video so they can be isolated or composited onto different footage.
Demucs: An open-source audio-separation model that splits a soundtrack into its vocal and instrumental stems.
Draft mode: A cheap, lower-resolution preview tier in Seedance 2.5 that lets you iterate on a generation repeatedly before paying once for the full-resolution render.
Face mesh: A tracking overlay that maps 478 points across a face so an AI model can reproduce exact facial motion without ever seeing the real face underneath.
Enhancor: A third-party API provider used in the video to access Seedance 2.5 more cheaply than going direct, without training on submitted data.

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

Hi, welcome to this masterclass. My name is Syrio and this is entirely AI. I'll show you.

So that's me in a scene that I was never actually in and I didn't even use any app to make it. And this is what actually happens when you try to do it the normal way. I'm uploading that same template video into Cdance 2 .5 and asking it to swap me in.

Blocked but apps like Hicks fields can jutsu do this exact thing all day every day with these exact videos So how the are they getting around it?

Hey friend, welcome back Today we're going to reverse engineer Higgs field scan jutsu and we're going to get around the IP blocks in C dense 2 .5 so that you guys can put yourself into pretty much any movie scene or talking head you want without any special software. A quick heads up before we start, this video is for educational purposes only.

I basically want you to understand how these AI models actually work and what they can really do. And this might end up being the first episode of a new series where I actually show you guys the stuff that they do not want you to know. And by they, you know who I'm talking about.

Okay, so no fluff. Here's how this is going to work. We're going to start by doing the whole thing by hand, okay?

It's actually going to break on us twice and we're gonna fix it both times. At every step along the way, I'm going to pause and tell you why it works and not just what to click. And then I'm gonna give you the AI, the automatic version, the agent version.

the workflow so that you can turn it into a template for your own stuff. And in the next videos, we're going a lot further and we're going to make things like this. I'm Kim Kardashian, and this is Masterclass.

I'm Serio, and this is your time to build something of your own. Cut. Good job.

So what do you think? Was it any good? Man, you killed it.

But first, the basics. So let's not waste any more time. You guys have all seen these videos that are basically everywhere right now.

And the way that it works is that you upload a photo of yourself and then you pick a template video and then it just swaps you in for the person that is in that template video. It's the same moves, the same everything, the same lip syncing. And most people get this from one of those wrapper apps that sells it as a feature.

And like I've said before, that's really the whole reason that a wrapper works and good for them. What I mean by that is that they are selling to you an api workflow it is not the ai model it's the workflow it's the thinking they are selling you the thinking that they do for you and you just press a button and so this tutorial is for those people who actually want to build that button i spent the last 12 to 24 hours on this going through open source models masking cutting people out and a whole bunch of other subs that didn't work and i ended up with the one that i'm going to show you today so Let's go back to that block generation from the start.

This is the template. And all I did here was I head over to CDANCE 2 .5. I uploaded this exact video and I typed, replace the character in the video with the picture that I uploaded.

And it got blocked for IP, which I mean, sure, it makes sense. These people are famous. But like I said, this exact video works totally fine in Genjutsu.

But before we actually build anything, I think that it is important for us to understand why. Because if you get this part, the rest of the video is basically just details. There are really only two things that get you blocked here.

The first one is the face. The second one is the music. And the way that I like to think about this is that a video is really carrying these two things at the same time.

It's carrying who's in it. So the face, the voice, the music. And it's carrying how things move.

So the motion, the timing, and the lips. And the filter, the IP filter, only cares about the who. It does not care about the how, and the how is the movement.

So the whole idea here is that you keep the how, the movement, and you throw away the who, the IP. And everything that we're going to do from here on out is really just hide one of those two things, obviously without losing the movement. So the first thing that we're going to have to do is turn this template into something where the AI cannot tell who the people are, but it can still...

See how they move and for that we're going to use something that is called depth anything you can find it on hugging phase or replicate or foul and All you have to do is that you upload your video you click generate and then you get something like this So this is what's called a depth mask and what I mean by this is that it throws away what everything looks like and it keeps how close or how far things are from the camera.

Like, it's the depth. If you've used Photoshop or you've done retouching before, it's kind of very similar to Dodge and Burn, but in this case, it tracks all the movements. The movements are still there, but the faces, as you can see, they're not.

It's just a silhouette. This one is from Fal, and this one is from Replicate. I like the one from Replicate.

I don't know. I just like the colors of it. There's no difference.

You can use the one from FAL too. And there's one more thing, which is the audio, right? If we want the lips to match the song, then the audio has to go through as well.

Now, the first thing that most people would say is just mute it. But we cannot do that. And the reason why is that AI actually needs to hear the voice in order for it to move the mouth.

So instead of muting it, we're going to keep the voice, but get rid of everything else. keep the actual words, remove the soundtrack, remove the background music. So what I'm going to do is that I'm going to take this depth mask and bring it into CapCut.

And I'm just going to make a new project and drag the clip in. And then I'm going to click on it. And over here on the right hand side, I'm going to go to audio, then voice changer, then voice filters, then I'm going to pick pitch up.

This is a voice filter and what it does is it just changes the pitch of the music. And then once that's done, I'm going back to the basics. I'm going to click on isolate voice, then keep only the vocals.

And so now what's left is basically only the acapella without the music or soundtrack. So listen to it.

So that's both of them hidden, right? C -Dense 2 .5 isn't going to block this anymore because there are no faces, there's no music, and we still have all of the movement, right? Pretty neat.

Now, instead of giving it the original video, we're going to give it the depth mask video plus our own photos and then swap the people in. But there is one problem though, which is that the depth mask has no idea what the place actually looks like. It just either like grayscale version of the depth mass like in foul or like this other one here And I think it's easier to show you guys So let's run a quick test in order to do that.

We need an API key So I'm going to grab my enhancer API key for C dance 2 .5 So we can test and the link for that is in the description And yes It is cheaper than all the other platforms that say that they are the cheapest and also doesn't train on your data So your stuff is safe I'm going to copy the docs right here, and then I'm going to paste them into Codex or Cloud.

I'm using Codex right now. That's really because I already had this project started in here. It does not mean that I'm endorsing OpenAI or anything.

It's just easier for me right now. Hey, can you set this up for me? Then I'm going to head over to my Enhancer API dashboard.

I'm going to generate a key, copy it, and... I'm going to go back to the agent and I'm going to ask it to make an ENV file, which is just like a little private file in your folder that holds your keys and saves the keys in there for safety. So please, guys, never, ever paste your keys in the chat.

Always put them in an ENV file. I'm going to click on here and I'm going to paste in my API key and we're all set and we are ready to generate. But before we actually make anything, the next thing I want to talk about is.

Draft mode because a lot of people get this wrong what I mean by draft mode is see that's 2 .5 lets you make a 480p version of your video first over and over Until you like what you see and then it finishes that exact same video in 1080p So you don't have to lose any detail and it's not like an upscaler It's kind of like doing a pencil sketch before you paint you can redo the sketch over and over again as many times as you want and then when you're ready you're like cool, I'm gonna pay for the paint once.

So use draft node every time that you're trying to draft something or generate something with Cdense. Now let's say that you get a draft that you really like. What a lot of people will do is they're going to download it and then upload it right back as the reference video thinking that's the same thing.

It is not. And the reason why is that the draft is tied to a task ID. and not the video file.

So when you upload that file again, that's actually a totally different kind of request, and it doesn't carry over any of your settings. So there's no promises that you're gonna get the same video back. So basically you do all of the cheap tries in draft mode, and then you only pay for 1080p one time.

Hopefully this makes sense. You're gonna see it in action. So I'm just going to drag in the depth mask, this one, and then the two photos that I wanna use.

which is me and Teen. And if you guys haven't met Teen yet, Teen is this little creature that I've made with AI, basically my companion, and also a really cool new project that I'm working on that is also linked in the description as well. And then I'm going to paste in this really simple replacement prompt.

You can screenshot that. And then I'm just basically going to say, hey, use the draft mode for this generation. cdens also needs links not files so please throw everything on a temporary server like cat box or temp files it's going to host your inputs temporarily and then you're going to hit enter and this usually takes about one minute okay cool so here's what we got and if you want to grab the video you can either get it from the agent or you can go into the logs and usage in your enhancer dashboard and give it a quick refresh and it will be sitting right there with your request id that's also going to match your agents And I mean, it's not great, right?

Well, we basically have two problems here. The first one is that it didn't actually swap me in all the way. So that's my face, but that it's very much the actor's body with all the long hair.

And the second thing is that the whole thing is just... Gray so one way to fix this is to have the agent describe the studio and the reason why is like I said earlier the depth mask has no idea what the Place looks like and if this was your reference video, this is what you got back So that's where the gray comes from and what I'm gonna do is I'm gonna take a screenshot of the original video and I'm gonna drop it into the agent and I'm gonna say Hey, describe this place for me, but do not describe the people and then update the prompt with that and send the same request again.

So that's the prompt that I'm using. You can pause and screenshot this. Since we're doing this with the agent, the agent knows exactly what they sent before and they can just update that request or update that payload.

You don't have to do anything manually because the agent has context of your workflow. And I'm just going to hit enter and submit it. Okay, now we have caller.

Which is nice That's one of our two problems fixed, but I mean the other problem is still there for some reason It is still sticking way too close to the depth mask, which is why It's still the actors body. It's not mine.

I'm not entirely sure why I think it's basically I don't know I think it just sees an outline and then it traces it but I've got one more trick to fix this So this time I'm going to grab the original video and then our depth mask with the changed audio I'm going to drop the original clip into CapCut and then I'm going to put in the depth mask right on top of it I'm gonna mute the original clip.

So the only audio left is our Pitch vocals. And I'm going to click on the depth mask and go over to the video tab, then the background removal, then click on custom removal and just highlight the people.

Basically what this is doing is that it's cutting out just the people and then it's laying them on top of the original video. Now, why does that work? Well, if you think about it, like now it's a very different video.

We were giving it a depth mask, which is really just the outline. So it was tracing the outline before and now we are giving it something that that is totally different and it stops from tracing it So, I don't know it looks at the whole thing again, and I'm gonna export that this is what it looks like then I'm gonna send it back to the agent and say hey swap the old source video for this new one and Make another draft and look at that That worked That's pretty cool.

Okay, so that was basically four moves. We hit the face, we hit the music, we described the room, and then we laid the people over the real video.

And that's way too much manual work to do every single time. But you can do all of it with one click the same way Hicksfield does with a workflow. So I had the agent turn everything we just did into one.

All you need is a Replicate account, and I'm giving you a skill that sets everything up. Hugging face is also optional, just a backup choice. So really quickly, it is 11 steps.

One, we upload the original video. We keep the original video for also the final audio. Two, Demacus on Replicate splits the vocals from the music.

Three, Rubberband pitches the vocals up by three semitones. Four, we have depth anything video. which makes the colored depth video that we talked about.

Five, we have Sam 3 cutting out all the people on every frame. Six, it puts those depth people over the original background. Seven, the pitched vocals go right inside of that video.

Eight, you check everything before it sends. Nine, C -Dance makes a draft through enhancer in the edit mode. 10, it throws away C -Dance's audio and puts in the whole original soundtrack back.

And 11... You approve it, and it makes a 1080p. Video and you download the final mp4.

I know well, but this workflow is making millions of dollars Now you don't have to do any of that yourself I'll show you all of it running as soon as it is set up So this whole thing runs through your agent and it comes with its own interface I built all of it for you guys All you have to do is download the skill give it to your agent and then plug in your api keys from enhancer And replicate so you can grab the skill in the description grab the folder and we're going to drag it in and and we're gonna say, please read the README file and set this up for me and just let the agent do it.

The agent is gonna set up everything for you and ask you some questions along the way. And if you don't wanna download the files for whatever reason, you can just ask your agent to build this workflow over here. So you can screenshot this or just give it my video transcript.

this YouTube video now I already have mine set up and There are two ways to run this the first one is you just drop in the video That you want to change and then the photos that you want to be placed in the video or like swapped And you say hey run the workflow as simple as that the second one you do it yourself in the UI interface And that's what I'm going to do here, so you guys can see every step so the interface is really just four steps because you don't see everything that's happening in the backend.

You have to upload, review, draft, and export or save. So I'm going to upload my video, which can be an MP4 or a MOV up to 30 seconds. I'm going to leave it on advanced because that's the depth workflow we just built.

And then I'm going to click prepare. And so right now it's doing steps through. 1 to 7 all by itself.

You don't have to do anything. It's splitting the vocals from the music. It's pitching them up.

Everything we did in CapCut. And then it's making the depth video, cutting the people with SAM 3, laying them over the original background, and cleaning up every little stray or edges that are off from that mask. This is step 8, and it's not going to let me generate until I've watched the whole thing and I've approved it myself.

And the reason for that is that my be a bad mask. So we don't, I don't want you to basically just throw money down the drain.

AI agents are unpredictable. That's why we make sure that they check their work every time. So it's checking for holes in the body, the hair or hands.

This one to me looks very clean. So I'm going to prove this and then I'm going to add my characters. So character one, that's me.

Character two is optional. In this case, it's teen. And then the prompt is already written.

And you can edit that if you wish. And I'm going to generate my draft. Cool.

So here's our draft and listen to that.

My 10 principles that build billion dollar brands. Take notes. That's the original audio, not the pitched one.

And I'm pretty happy with this, so I'm going to approve that. And then it makes the 1080p and puts the original audio back one more time. And that one does cost extra and it's never gonna do it without asking you first or approving now that workflow is the best one if you want to bulletproof your Generations and it mostly uses the audio to match the words So if your video has no audio, it will still generate for sure But there's one more workflow for when you want the face to move exactly like the original and it's the same workflow, but we add what's called a face mash and what i mean by that it's a little wireframe that gets drawn right on top of the face it's kind of like those dots that they put on actors faces in the behind the scene video so they can track them so the dots can capture how the face moves and then the computer can Totally create a different character on top of that.

So that same thing same thing applies here It's gonna track how the face moves and it's still gonna hide the person behind it And this is actually the easiest one because it only has two steps so you can change the pitch you can add the face mash and there's no Cutting anyone out like that like it just removes all those extra other steps.

So if you guys Downloaded the project. You can just ask the agent to use this workflow instead It's faster and it's also right here in the interface.

It's called Fast or you can ask it to install the face mash tracking and put it on top of the original video You send that in and you get something like this. That's dope. That's crazy.

So what the face mash is doing is it's tracking 478 points on the face So around the eyes, the lips, the nose, the whole face. And it is drawing that mesh on every single frame.

And it saves all of those points with all the timing in a file that we have set up to be called facelandmarks .json. Everything is done and safe for you. You don't have to worry about it.

And it uses Google's face landmark model. from Google's official model storage. One heads up though, at least right now, this works best with one person, which is centered, talking to the camera.

If you have two people moving around, I would just stick with a depth mask workflow. And if you want to bulletproof this even more, you can ask the agent to put the face mash on top of the cutout depth video. So on top of this one right here, and then you get something like this, and then you send that in as a source video.

And again, all of this works by just talking to it. There's nothing for you to install yourself. You just ask the agent to set up the Google MediaPipe Face Landmarker or any other packages.

Okay, so this has been a long video. So from here, you guys can keep making things better. You can add a transcriber.

So it writes down what's said in the video and puts that into the prompt. You could have the agent watch the original video and describe the movement in detail second by second. And then you can even use your own voice.

So you basically make your own avatar. Look at this. And I'm here to spill the secrets behind APIs so that you can build million dollar workflows.

That is it. And if you take one thing away from this video is that. Everything, every one of these tricks was basically just the same idea implemented in different ways and methods.

You just keep how things move and you throw away who is in the video. And hopefully you guys know how Genjutsu actually works now and how to build it yourself. I have attached the entire open source project in the link in the description.

So just get it from there. It's free and go make something with it. And in the next one, we're going to go...

way past the basics. See you guys there. And do not forget, create without limits.

This is Syria.

The Hook

The bait, then the rug-pull.

He opens with a slick AI-generated 'masterclass' clip of himself, then shows the exact same request getting blocked when submitted directly to Seedance 2.5, setting up the question the rest of the video answers: how does Higgsfield get around its own underlying model's IP filter?