How to keep the same character in every AI video shot: face, outfit and voice

You generate a great first shot. The second shot has a slightly different nose, the jacket changed color, and when she finally speaks she sounds like someone else. It's one of the most common problems in AI video, and it gets worse the more shots your story has.

This guide covers the three things that have to stay the same for a viewer to believe it's one person: the face, the outfit and the voice. Most guides stop at the face. The voice is where a lot of AI short films fall apart, so it gets its own part here.

Every step is done on the Sumi Studio canvas, and most steps come with a prompt you can copy.

How to read this

If you're trying this for the first time, do a small test: one character, two shots, one line of dialogue. Follow parts 2, 3 and 4 with that, and you'll see how the whole process works.

If you're planning a full short film or a series, read everything in order. Part 5 lists the shots that break consistency most often, which saves you a lot of rerolls.

Here's what each piece on the canvas does:

Part 1: Why your character keeps changing

Every generation starts from zero. The model doesn't remember the last clip you made. All it has is what you give it in that one request: the prompt and whatever reference images or audio you attach.

So when the only reference is text, the model reads "a woman in her 30s with short dark hair" a little differently each time. Small differences add up over ten shots, and by the end it's a different person.

The fix is to give the model the same visual reference and the same voice reference in every shot, and write the description the same way every time. On the canvas, that means one character, saved once, connected to every shot.

Part 2: Lock the face and the outfit

2.1 Make the character on the canvas

Open a new canvas and hold Tab to bring up the node wheel. Pick Image to add an Image node.

Write the character in the prompt box below the node, in the same art style you want for the final film. Describe everything that must stay fixed: face shape, hair, skin, build, and every piece of clothing down to the shoes. Ask for a full-body image on a plain background, so nothing in the scene distracts from the character.

Write the position of any small detail. "A small scar above the left eyebrow" or "a silver ring on the right hand" stays put. "A scar" and "a ring" will move around between shots.

If the result isn't right, regenerate the node. Each node keeps its history, so you can go back to an earlier try if a later one comes out worse.

2.2 Turn it into a character sheet

A single image isn't enough. The model needs to see the character from several angles, or it will invent the back of the head, the side profile and the shoes on its own.

Click the plus on the side of your character node and drag it out to a new Image node. The character image now goes into the new node as its reference. Paste the prompt below into the new node. It works for any character, so you can reuse it as it is.

Create a character reference sheet of the character in the reference image. Keep the character's design exactly as shown: same face, hairstyle, hair color, eye color, skin tone, body proportions, outfit, accessories and colors. Do not redesign, restyle or add anything.

Layout: one single horizontal row of four full-body views on the left, side by side, never stacked on top of each other. The large face close-up sits on the right side of the sheet. Every full-body view has the same height and runs from the top of the head to the bottom of the feet, with empty gray space above the head and below the feet.

1. Full-body front view: facing the camera, standing straight, arms slightly away from the body, showing the face, full outfit and accessories.
2. Full-body side view: exact 90 degree profile, standing straight, showing the nose, jawline, hair shape and outfit from the side.
3. Full-body 3/4 view: body and face turned diagonally, about 45 degrees from the camera, standing straight.
4. Full-body back view: standing straight, showing the hairstyle, the back of the outfit, the back of the legs and both feet.
5. Large face close-up: head and shoulders, facing the camera, neutral expression, sharp and detailed.

The character has the same face in every view, with the full head visible in all poses. All four full-body views show the complete body including knees, calves, ankles and feet, nothing cropped, same scale, same standing pose and same neutral expression. Plain medium gray studio background, soft even studio lighting, no props, no text, no labels, no extra characters.

In the node's parameters, set the aspect ratio to 21:9 if the model offers it, otherwise 16:9. A wide canvas keeps the four full-body views from getting cropped.

Check the sheet before you move on. Every view should show the same face, the same hairstyle and the same clothes. If one view is off, regenerate the node now, because a mistake on the sheet gets copied into every shot that uses it.

2.3 Save the character as an element

An element is how Sumi Studio remembers a character. It holds a name, a short handle, reference images and a voice. When you type @ and the handle in any node's prompt, the images are attached to that generation as references automatically.

The quickest way is right on the canvas:

1. Right-click the character sheet node and choose Create element. The New element window opens with the sheet already added as a reference image.
2. Set the category to Character and give it a name and a handle, for example @mara.
3. In the description, write the fixed details in a few short phrases. The description is sent along with the name every time you generate, so keep it to things that never change: "short black bob, small scar above the left eyebrow, olive green bomber jacket, white sneakers".
4. Click Create.

You can also create it from Assets, under Elements. Click New element and upload the sheet with Add reference image. You can add the face close-up as a separate image too.

If you use Kling O3 or later, you can also add one short video of the character (MP4 or MOV, 3 to 8 seconds, 1920×1080 or 1080×1920). Kling uses it as a video subject. Other models only use the images.

2.4 One outfit per element

If your character changes clothes in the story, make a second element for the second outfit, for example @mara and @mara-night. On the canvas, connect the character sheet to a new Image node, ask for the same character in the new outfit, then make a sheet of that and save it as the second element. Keep the face close-up the same in both.

Putting two outfits in one element gives the model two answers to the same question, and you'll get a mix of both.

Before you move on, check these:

• Every view on the sheet shows the same face and clothes
• Small details have a written position (left or right, top or bottom)
• The element description only lists things that never change
• Each outfit has its own element

Part 3: Lock the voice

A character can look perfect in every shot and still feel wrong if the voice changes. Video models that generate dialogue pick a voice on their own each time unless you give them one to follow.

In Sumi Studio, the voice is part of the element. You can make the voice on the canvas and then attach it to the element.

3.1 Clone a voice on the canvas

Click the plus in the left menu, go to Audio and choose Voice. This adds a Voice node (it shows as Text to speech on the canvas). In the bar below the node, click the current voice name, next to the model. This opens Select voice, with the voice library, your own voices, favorites and the voices of your elements.

Click Clone voice. You read a short passage aloud for 10 to 20 seconds, or upload a recording.

The clone screen shows the recording requirements for the voice model. For Qwen Audio 3.1 TTS Plus, the recording needs at least 5 seconds of continuous speech, pauses under 2 seconds, and no music, singing or noise, in a WAV, MP3 or M4A file up to 10 MB and 60 seconds. Only clone your own voice or a voice you have permission to use.

3.2 Or design one from a description

If the character doesn't sound like anyone you know, design the voice from a description. Open Select voice from the Voice node the same way, then click Design voice and fill in the form:

1. Voice name: use the character's name, for example Mara, so you can find it later.
2. Model: pick the voice model. The voice fee is charged once, the first time you generate speech with that voice.
3. Preview text: the line the model reads so you can hear the voice. It's required, takes up to 500 characters and is billed by length. Write something your character would actually say, in the tone you want.
4. Voice description: describe the voice by age, tone and pace. Avoid vague words like "cool" or "nice", which the model can read many ways.
5. Language and Gender: set both to match the character.
6. Click Create voice.

I told you I'd come back, and here I am. It's been a long time since I walked these streets at night. The rain hasn't changed, the neon signs haven't changed, and you still leave the back door open. Don't look so surprised. I'm not here to make trouble. I just need one answer, and then I'll be gone before the sun comes up. So sit down, pour yourself a drink, and tell me the truth this time. Who sold us out that night? Take your time. I've waited three years, I can wait a little longer.

A woman in her early 30s, low and slightly husky voice, calm and measured pace, speaks quietly but clearly, a little dry and sarcastic.

The new voice shows up under My voices in Select voice.

3.3 Attach the voice to the character

Go back to the element in Elements. There are two places to set the voice:

• Under Voice, pick the voice you cloned or designed from My cloned voice.
• Under Character voice reference, choose or upload a clip between 2 and 10 seconds long. Seedance, Wan and Hailuo use this clip when you @ the character.

A good reference clip has one person speaking clearly, no music and no background noise, in the tone you want most of the time. If your character is mostly calm, don't use a clip of them shouting.

Kling only uses voices cloned with Kling. If you want your character to keep the same voice in Kling videos, choose a Kling-cloned voice under Voice.

One more thing: when a node on the canvas picks up an element's voice, it keeps a copy. If you change the voice in the element later, nodes you already made keep the old one. Settle the voice before you build your shots.

Part 4: Use the character in every shot

4.1 Always call the character with @

Hold Tab and add a Video node for each shot. In the prompt box below the node, type @ and pick your character. That's what attaches the sheet and the voice. A shot where you describe the character in words but forget the @ is the shot where the face drifts.

A new Video node starts in Text to video mode, which doesn't use your character's references. Switch it before you generate: click the settings bar below the prompt (it reads something like "Text to video · 16:9 · 480P · 8s") to open Parameters, and under Generation method choose Omni reference. In this mode the model uses both the character's images and voice.

4.2 Write the shot, and leave the look to the element

The shot prompt should describe what happens: action, camera, place, light and dialogue. You don't need to describe the face again, and rewriting it in different words every shot can pull the model away from the reference.

Keep the background short but specific. Without one, many models give you a plain, empty set.

8-second shot.
@mara walks toward the camera along a narrow street at night, wet pavement with neon reflections, a few closed shops in the background.
Medium shot, eye level, slow push-in.
She stops, looks up and says quietly: "I told you I'd come back."
Calm, low voice. Only the sound of rain and her footsteps. No music, no subtitles.

Before you hit generate, the node shows how many credits the shot will take.

4.3 For the shots that matter most, start from a first frame

For a close-up or a key moment, make the first frame as an image first. Add an Image node, @ the character and describe the composition of the shot. When the frame looks right, click the plus on its side, drag it to a Video node, open Parameters from the settings bar and set Generation method to Image to video, or First/Last Frame if you also made the last frame.

This gives the video model a fixed starting point with the right face already in place, which is usually the most reliable way to keep a character consistent. It also costs less to get right, because retrying an image is cheaper than retrying a video.

Small things that tend to disappear during a clip, like a hat, a tattoo or glasses, should be visible in the first frame. If they're only in the text, they often vanish halfway through.

Here's an example with Mara in her night outfit.

@mara-night in a close-up, standing under a flickering streetlight in a narrow alley at night, light rain falling, wet brick wall behind her slightly out of focus. Warm orange streetlight from the left, cool blue neon glow from the right. Her coat collar is turned up and she looks just past the camera with a calm, tired expression. Her face sits on the right third of the frame. Eye level, 50mm lens, shallow depth of field. Photorealistic cinematic, natural skin texture, 35mm film look. No text, no watermark.

5-second shot.
Use the reference image as the first frame and keep its composition, lighting and framing exactly the same. @mara-night stays under the streetlight in the rain. She slowly lifts her eyes to the camera, lets out a short breath and the corner of her mouth tightens. She does not speak.
Close-up, eye level, very slow push-in.
Rain falls steadily, the streetlight flickers once, a car's headlights sweep across the brick wall behind her.
Only the sound of rain and a car passing in the distance. No music, no subtitles.

4.4 Build all your shots from one reference

You don't have to upload the character sheet again for every shot. Drag a connection from the sheet node to every Video node that needs it. Whatever is in the first node goes into the next one, so one reference can feed as many shots as you need.

If one shot comes out wrong, regenerate only that node. The rest of the canvas stays where it was. If you want all the project's images and clips in front of you while you work, drop a Stack node on the canvas.

When the shots are ready, open Smart Cut. Everything you generated on the canvas is already there to put on the timeline.

Before you move on, check these:

• Every shot prompt with the character has the @handle
• Shot prompts describe action, camera and place, and leave out the face
• Key shots start from an approved first frame
• Hats, glasses and tattoos are visible in the first frame

Part 5: The shots that break consistency

Some shots are much harder for video models than others. If you plan around them while writing the story, you'll reroll a lot less.

Common mistakes

• Starting with video before the character sheet is approved (2.2)
• Writing small details without a position, so they move between shots (2.1)
• Two outfits in one element (2.4)
• No voice on the element, so the model picks a new voice every shot (Part 3)
• Changing the element's voice after the shots are built (3.3)
• Forgetting the @handle in a shot (4.1)
• Describing the face again in every shot prompt (4.2)
• Generating key close-ups straight from text (4.3)
• Wide shots and turn-arounds where the face matters most (Part 5)

Set up your first character and start with a two-shot test before you build the whole film. To learn more about the canvas, elements and the models, see Sumi Studio Academy.