Speaker A
I NEED THIS DOG'S HEAD ON MY TABLE. [music] HEY, I'm Adam, the star of this film, and today I'm showing you how this film was made. How to make impressive fight scene, how to create handsome characters like me. And
A creator walks through making a short AI noir film in Higgsfield, covering character design, video voice elements, locations, prompt skills, fight choreography and acting refinement.
Key Takeaways
- Video-based character elements with recorded audio produce far more believable acting than image-only references.
- Structured project organization and named assets make prompts reusable and reduce wasted credits.
- Using AI to expand brief directions into detailed shot-by-shot prompts, combined with specialized skills, significantly raises output quality.
- Drafting at lower resolution before final renders is an effective way to control generation costs.
What the video covers
- The video is a full workflow guide for making an AI-generated short film with Higgsfield AI using Seedance-style tools and Claude for prompts.
- Characters are developed by generating model sheets from a reference photo, then cleaning artifacts by adding texture references for hair, skin and fabric.
- Image elements are contrasted with video elements; video elements (with recorded voice audio) produce more natural facial motion and acting.
- Voice references are simple recordings made by the creator or friends, used to give AI characters a human tone without copying anyone's voice.
- Locations are generated with specified lens characteristics (24mm/35mm anamorphic looks) and expanded into detailed prompts via Claude.
- A structured scene prompt is built by naming assets as elements, using a Claude prompt skill, and specifying shots, camera moves, dialogue and timing.
- A cost-saving tip is to run drafts at 480p to check framing and blocking, then regenerate only the winning takes at 1080p.
- Fight scenes are improved with a fight-choreography skill that defines how hits land and how opponents react, plus frame-by-frame direction.
- Acting is refined with an acting skill that adds performance direction such as pauses and emotional beats to dialogue.
- Common mistakes are highlighted, such as inconsistent object details (e.g., a bullet changing shape) and generations that place characters incorrectly.
Chapters
- 00:00Introduction and Film Teaser
- 00:58Project Setup and Character Design
- 02:46Cleaning Up Character Artifacts
- 04:06Casting Antagonists and Video Voice Elements
- 05:39Building Locations with Lens Looks
- 06:41Generating the First Scene
- 09:07Fight Choreography and Camera Fixes
- 10:50Acting Refinement and Continuity Tips
Full Transcript — Download SRT & Markdown
Speaker A
what are you doing? What you [music] doing here? I thought you left. Left? This is my channel. Slide over.
Speaker A
Oh, come on. I was just setting it up for you. Oh, sure you were. Oh, go hunt something. Anyways, sorry about him. I made the short film in just 2 days with Huxfield. I'll show you how to choreograph fights, nail the acting, and
Speaker A
direct the camera with the exact prompts, skills, and steps I used. Steal this workflow for your next film. I'll show you which mistakes cost me credits, so you can skip them. All right, wait.
Speaker A
They haven't even seen the full film. [music] I know they will at the end.
Speaker A
Anyways, let's break it down. First, I'll focus on the parts you can use in your own films. Every prompt and asset, including the shots I skip, is in the blog post below. First, I create a project in Cinema Studio. One folder per
Speaker A
scene with subfolders for characters, locations, and props. A little setup now saves a lot of scrolling later. I'm using skills for scene structure, fights, and acting. [music] I'll show you what each one changes as we go. Now, the cast. Adam started as me. White
Speaker A
[music] hair, half vampire, hunts werewolves. I give Clark my photo and a few notes. pale skin, icy blue eyes, the long leather coat, and ask for two versions: curly hair, and straight.
Speaker A
Claude expands that into detailed prompts. I usually use soul for my characters, but it doesn't support image to image, so instead I'm using GPT image 2.5 for this trick. I use the same photo with each prompt and generate two
Speaker A
versions. Here's what I got. Both look pretty good. I'm going with a sharper face and straighter hair, but let's look a little closer. So, the skin has these weird spots and patches, and the hair looks damaged. [music] My hair is much
Speaker A
healthier than that. And I'm not happy with the shirt's texture either. Watch what happens when I try to animate it.
Speaker A
Yeah, see that super dirty plastic skin, dead looking [music] hair, and tons of weird GPT artifacts moving around. It just looks cheap. Here's how I clean that up. For this, I'm using Cream. I just drop in Adam's model sheet, then
Speaker A
add reference shots for his hair, skin, fabric, and the [music] background. You can easily pull these online. Then I specify what to pull from each image.
Speaker A
So, [music] hair strands from one, skin detail from another, and what to keep consistent. Think of it like texturing a 3D character. I'm giving the surfaces more realistic detail. These reference photos show the model exactly what I want. Here's [music] how it turned out.
Speaker A
There we go. Now I have a clean version of the character. Let's compare both generations side [music] by side. Yeah, there we go. Much better. Pores, not patches, and the hair moves well, like actual hair. Now, in just a minute, I'm
Speaker A
turning this sheet into a fully reusable video reference, complete with his custom voice. This is how I look, and this is how I sound. But first, let's finish the cast. Adam needs someone to fight. For Victor, the antagonist, I'm
Speaker A
using AI cast. I set the budget, era, and archetype. Then, fine-tune his features. It's a great shortcut for when you know the type, but not the exact look. Generate a quick batch and see who lands. I'm going mid-40s, tall with gray
Speaker A
hair, a beard, and formal clothes. Out of these options, I'm going to go with this guy. He fits the story's vibe perfectly. For the boss, Ron, I asked Claude for a heavy set gangster boss and gave it some details on his look. And
Speaker A
I'm generating a bunch of options using the cheapest yet the most creative model, Soul Cinema. Out of all these, I'm picking this one. Now, watch these two clips. Which one feels more real?
Speaker A
Where were you? I called you 11 times. Where were you? I called you 11 times.
Speaker A
So, this one uses an image element, and this one uses a video element. An image shows the model what the character [music] looks like. A video also shows how their face moves and how they sound, and the acting feels much more natural.
Speaker A
So, instead of relying on image elements and burning credits on takes where the face looks off, I'm creating video elements for all three characters before I build the scenes. This is how I look and this is how I sound.
Speaker A
This is how I look and this is how I sound. This is how I look and this is how I sound.
Speaker A
For all characters, I use simple recordings like these ones. I know they sound quite silly, but human tone makes the AI voices feel more alive without trying to copy anyone's voice.
Speaker A
Exactly. For Adam, I recorded the sound myself. For Ron and Victor, I asked my friends to record theirs. Now, here's how I make the video element. Select your character generation. Create an element and choose video. Now upload the
Speaker A
audio. Add a short prompt for how the character should move and speak and hit generate. Now locations. First a dark alley with an old phone booth. I ask for a 3/4 view to give the space depth and a
Speaker A
24mm anamorphic look, wide framing, softer edges, and oval bouquet. Claude expands that into a full prompt. And then I paste it into Soul Cinema and generate a few options. And I went with this one. Unlike the others, it's a
Speaker A
little brighter in the corners and the booth [music] is nicely lit. Next, Ron's office. Dark wood, a big desk, and a leather chair. Same quite a view, but this time I'm using a 35mm anamorphic look for a tighter frame. Oh yeah,
Speaker A
that's the depth I wanted. A solid [music] noir mood with enough shadow detail to read the room. To avoid typing all those lens details every time, I used an anamorphic scale in Claude.
Speaker A
Here's the difference in motion. Look at the background here. So in this take it stays sharper but in this one it's softer so Adam stands out way more. So that look starts in the reference and carries into the video. All right let's
Speaker A
generate the first scene. This goddamn [music] puppy need to learn its own place. I knew the things would go south. Now they killed Kenny.
Speaker A
I NEED THIS DOG'S HEAD ON MY TABLE. This piece of [ __ ] the phone call. Here's how I built it. I name and save each asset as an element in Hfield. Then give Claude the same file a name. Claude uses that tag in the
Speaker A
prompt. So Hixel attaches the right element when I paste it in. And I'm using the Hixel prompt skill, which is a set of instructions for Claude. I'm attaching it with my scene brief and asking Claude to use it to build a
Speaker A
conflict. [music] Then I ask for shots, camera moves, and dialogue. The skill still needs a direction, so I spell out how Ron's anger builds, which lines he shouts, and where the camera goes. The more clearly I describe the scene I have
Speaker A
in mind, the more specific Claude can make the prompt. Let's look at what comes back. Okay, these are the elements I named earlier. Then there's objects for framing, camera for angles and movement, and action timing for [music] second by second facing, lighting,
Speaker A
sound, and acting help guide Seance. Then I check that into CES 2.5 with the references attached. 30 seconds 21 by9 and run a few batches to get different options. A little pro tip, this one saves real credits. Run drafts at 480p.
Speaker A
You're only checking framing and who's standing where. Pick [music] the winners, then recreate just those at 1080p. Okay, let's go through the batches. So, this first generation gave me Adam's starting shot. [music] The camera rises behind him and the 3/4
Speaker A
angle reveals the noir world. The other batches kept messing up Adam's position. He either ended up behind the booth or stuck in a flat frontal shot. Take two gives me Ron's outburst. Keep that performance in mind. It matters later.
Speaker A
Take three, the hang out cuts him off mid word. But look at the last batch.
Speaker A
The moment Adam gets on the bike and leaves. I didn't specify Dutch angle for the shot, but this take gave me a low tilted view that adds tension. This is why batching matters. The exact same prompt gives slightly different results,
Speaker A
and sometimes a [music] random take gives you a killer shot. So, I kept it.
Speaker A
Then, I recreate the selected takes at 1080p, check the new versions, and bring the approved footage into the edit. By the way, if you want feedback on your shot, post it with a prompt in our Discord. Somebody will definitely help
Speaker A
you out. I'm going to skip the B-roll breakdown. The prompts are in the blog post below. Now, onto the club scene.
Speaker A
Let's watch it first. [music] [music] Heat. [music] Heat. [music] [music] [music] I made this club location in Soul Cinema. Here's the prompt I used. I build the club scene in three parts.
Speaker A
Entrance, party, and the fight. Then I cut the best shots together. Block one, Adam walks in. start overhead, then follow him through the crowd in smooth slow motion. On the first take, no top down shot, [music] and the crowd's
Speaker A
movement felt off. So, I asked for two fixes. A high angle from the side and a crowd that moves naturally, jumping, landing, and making room for Adam. Real speed this time. Okay, two shots worth keeping. Adam's entrance works, but
Speaker A
right now, it's just him in the club. Adam's here to hunt. Everyone else is here to party. These shots give me that contrast. I'm taking two shots from this generation. the woman dancing, facing the camera, and these two guys enjoying
Speaker A
their drinks. Every extra needs a job. Different actions make the crowd feel alive. I also added a couple shots of Adam moving through the crowd and turning toward the camera. Then the guards spot him. One gives the order.
Speaker A
Another switches on his baton. Kill him. [music] Okay, the guards move in sync and the baton looks fake. So, I'm gonna give each one separate reactions and tone down the electrical effect.
Speaker A
Kill him. Oh, yeah. Much better. Now, the fight. Here are my first takes without the fighting skill. Right there. See that?
Speaker A
The punches feel soft and Adam looks stiff. I need sharper strikes and reactions that sell the impact. So, I use a skill for fights. It's packed with real fight choreography. how a hit lands, how the other guy reacts. I just
Speaker A
write the fight as a list. Who's where, what Adam [music] does, and how each guard goes down. So, medium shots show the move, and the close-ups makes the contact feel real. The skill holds fight physics really well. [music] Look at the
Speaker A
slow-mo. Dynamic hits, reactive camera work, and the club lighting sells every punch. Entrance, party, fight. Scene three, more guards, more fighting, and Adam is just getting started.
Speaker A
[music] Here's the corridor for the location. And here are the Yakuza crew. I generate the extras together as a group, so I only need one reference asset instead of creating a separate one for each character. For this fight, I wanted the
Speaker A
camera right inside the action, shaking and whipping between Adam and the guards. Without the fight scale, Adam's stance looks awkward, the heads feel weak, and the fight falls flat. So, I'm using the same fight scale, but this time, I'm spelling out the movement,
Speaker A
where the foot goes, where it lands, and which way the guards fall. Here's what I kept. First take, the guard's conversation with the acting skill.
Speaker A
Second, Adam's entrance and the start of the fight. I mean, watch that whip band.
Speaker A
The fast sideways movement gives me a place to hide the cut and keep the action flowing. The third picks up the fight and takes out the last two guards.
Speaker A
And the final gives me the ending. Four takes, one fight. You can't see the scenes because every cut is on movement.
Speaker A
And next up, Victor's office. Two guys with one gun. Now, the acting has to carry the scene.
Speaker A
Oh, Adam, I've been waiting for you. Sit down. Have a drink. First, the office. The first badge came out cold and blue. I like the room, but the mood wasn't quite right, so I didn't reroll. I just re lit it. Yeah, that's
Speaker A
much better. Now, it's a dialogue scene, so I'll be cutting back and forth a lot.
Speaker A
The room needs a second side. The same room, just facing the door with a hallway behind it. I'm also asking for damage by the handle because Adam's about to kick it in. Now, the sofa's behind Victor, the door's behind Adam,
Speaker A
and the model always knows which way it's looking. Two rules for dialogue. Lines go in word for word where the model rewrites a script. And don't plan every shot. Give it camera rules and let the skill build the shot list.
Speaker A
Oh, Adam, I've been waiting for you. Sit down, have a drink. Kenny's gone. I want this as a duel. Whoever loses his cool first loses. But right now, both guys stay calm. Victor needs to crack and watch this glass. This breaks the shot.
Speaker A
So, I go back to Claude and ask you to revise the prompt using the acting skill. But before I generate another badge, let me show you what the skill does. I tested it separately with Ron and Adam.
Speaker A
When you were powerless, I helped you. And this is how you want to thank me.
Speaker A
When you were powerless, I helped you. And this is how you want to thank me.
Speaker A
The skill is called Hixel acting skill. I know, so creative. It builds the performance around what each character is trying to get from the other person.
Speaker A
I mean, look at the difference in his performance. Now, let's use it with Victor. I switch on the skill and tell it who these two are and what each one wants. Plus three fixes. Adam's line gets shorter. No glass and room for
Speaker A
pauses. There's a line, Victor. You knew exactly where it was. Overall, this works pretty well. The acting is significantly better now. It's actually starting to feel like a real dialogue. I've got some usable shots here, but I still want to tweak a few
Speaker A
things. For the next pass, I'm sending Claude these three fixes. First, drop the dust and splinters from the door kick. Second, move around Adam so I can stay on his face during the line. Third, lock the room, the door behind Adam and
Speaker A
sofa behind Victor. I built the finished dialogue from six generations, keeping the useful moments from different batches. And now to the next part. Adam pulls the trigger. Victor doesn't stay down. I ask Claw to continue the scene for another 30 seconds. Adam shoots
Speaker A
Victor, but that doesn't stop him. And for the transformation, I want to try something with the camera. Oh, and a quick editing tip. I'm repeating Victor's last line from the last generation in this one. That gives me a
Speaker A
little overlap so I can choose where to cut from one generation to the next.
Speaker A
You're one of us, kid. Always were. Shut your mouth. The sequence is there, but look at the bullet. That looks more like a metal ball, and it keeps changing. Different shape, different hole. And in this one, it goes back into his head. So, this is
Speaker A
where I add two specific references. One shows the used bullet I want. The other is a photo of Victor with a wound at the right size. I'm also making the order explicit. The bullet comes out, it falls, and the wound stays visible until
Speaker A
the camera hides Victor. And now I have a bullet shot I can use. You can follow it out of his forehead and down to the floor. For the transformation, I used Adam's back to hide the change.
Speaker A
We should be on the same side, you and me. Okay, last part, the fight. Two sides of a room is great for dialogue, but for this specific fight, it confused the model. So, I shrink the stage.
Speaker A
Everything happens within 2 m of the red sofa. I also need to pin down where Adam and Victor start. Here's a useful trick.
Speaker A
I turn one location image into a short room tour in Cance. That gives me new views of the same space from different sides. Then I grab the overhead frame and mark where Adam and Victor start.
Speaker A
Victor by the sofa, Adam across from him. And now I'm attaching this in both Claude and Seance. I'm specifying that it's a blocking reference. Use it for their starting positions. Then follow the camera directions in the prompt.
Speaker A
This take gives me the close-up impact I asked for. Now I need wider angles at real speed so the action is easier to follow. So here's the trick. I'm taking this one shot of me and using it to
Speaker A
generate a different angle or this one or even this all from the same video reference. Pretty cool, right? For the fight, I'm uploading the previous take to Claude and asking it to follow the same choreography from different camera
Speaker A
positions. Now I feed that detail prompt with a video reference straight into Kixold. And look at the result. Same video, but now I've got wider views, low angles, and troughs through the furniture. I look for moments that match
Speaker A
the movement in the original take and cut those in. But that's not all. This take ends before the movie is finished.
Speaker A
Watch how I keep it going. I'm using extend in Cinema Studio. It's useful when I like the shot and just need a few more seconds to finish it. Here's the prompt and reference I used for the original take. Now I select forward
Speaker A
extend and paste in the prompt for what happens next. Adam finishes the move. Then the camera pushes in on his face.
Speaker A
Let's see how it flows. [groaning] [groaning] And after all that, the last scene was just one prompt. You'll find that in the blog post, too. Now, let's watch the whole film again.
Speaker A
This goddamn puppy need to learn its own place. I knew the things would go south. Now they killed Kenny.
Speaker A
I NEED THIS DOG'S HEAD ON MY TABLE. This piece of [ __ ] Where you go? Where you go?
Speaker A
Kill him. [music] Oh, Adam, I've been waiting for you. Sit down. Have a drink.
Speaker A
Kenny's gone. There's a line, Victor. You knew exactly where it was. You stepped over it anyway. Now it's time to pay. [laughter] Still running errands for those sad little people.
Speaker A
You're one of us, kid. Always were. Shut your mouth. If it were that easy, it wouldn't be any fun at all. Adam, we should be on the same side.
Speaker A
You and me. I'm allergic to dogs. Heat. Heat. [music] [music] [music] [music] [sighs] [clears throat] Adam, now every dog knows his place.
Speaker A
Hey. Hey, I'm done, Ron. I'm out. Out? What do you mean out? Your father's back.
Speaker A
Try this workflow and these tricks in your own film, then share it in our Discord. I'd love to see what you make and break down some of your films in a future episode. All right, [music] Adil, slide over. You have the intro. Let me
Speaker A
finish.
Topics:AI filmmakingHiggsfield AISeedanceAI short filmvideo elementsprompt engineeringAI actingfight choreography AIClaude promptsAI cinematography
