Consistent AI characters: keeping the same face from one image and one video to the next

The problem everyone hits after ten minutes

You generate a heroine for your project. She's perfect: thirty years old, short red hair, a faded denim jacket. You run another generation to see her in profile. And then, surprise. The hair has turned brown, the jacket is gone, the face has nothing in common with the first one. Same prompt, word for word. Completely different character.

This is not a bug in Imaginode, or in any other tool for that matter. It's the normal behavior of every current image model. And until you understand why, you'll burn hours and dozens of credits rerunning generations and hoping for a miracle. This article explains the cause, then the real solution, the one that actually holds up in production for a comic, a series of posts, or a short film.

Honest spoiler: the solution is not a better prompt. It's showing your character to the model instead of describing it, using reference images. On Imaginode, that happens through the Reference node, and we're going to build it step by step, from the master portrait to the green @mention, with the exact price in credits for every stage along the way.

Why the face changes: random noise

An image model doesn't draw the way an illustrator does, holding a character in mind. Every generation starts from an image of random noise, a soup of pixels drawn by chance, which the model progressively denoises while leaning on your prompt. Two different noise draws give two different starting points, and therefore two different faces at the finish line.

Your prompt only ever describes a category of people. "Thirty-year-old redheaded woman in a denim jacket" matches millions of possible faces. The model picks one that fits the description, a fresh one every single time. It never picks the same face twice, because nothing forces it to.

There's a second piece of bad news to add: the model has no memory. Generation number 12 knows nothing about the 11 that came before it. It doesn't "remember" your heroine, because it never met her in the first place. Every click on Generate starts from zero, with new noise and no shared history at all.

Why the ultra-detailed prompt is never enough

The first reflex, and we've all had it: describe more. Four hundred words on the shape of the nose, the spacing of the eyes, the strand of hair falling on the left. It does improve the stability of the easy elements a little, the hair color or the clothing. But the face itself keeps on changing.

The reason is frustrating but logical: language is simply too poor to describe a face. Try describing your best friend in writing, then hand the text to an artist who has never seen them. The portrait will be plausible, never a true likeness. A human face is thousands of micro-proportions that no vocabulary can capture.

And the longer the prompt gets, the more the model has to arbitrate between instructions that step on each other. Past a certain point, every extra word about the face is paid for in attention lost on the scene, the light, or the action. You end up with a stiff, frozen portrait instead of the living image you actually wanted.

The real solution: show instead of describe

What an artist would do in real life is ask for a photo. Recent models can do exactly the same thing: you attach one or several reference images to the generation, and the model uses them as a direct visual source instead of interpreting text and guessing at what you meant.

The difference is dramatic. Where a 400-word prompt gives you an approximate lookalike one time out of five, three good reference photos give you a recognizable character almost every single time. It's not magic: the model copies real proportions instead of inventing them from scratch on every run.

One warning, though: not every model can do this. On Imaginode, the references port simply doesn't appear on a node when the selected model doesn't support them. If you don't see the port, switch models instead of hunting for some hidden setting: there isn't one.

The Reference node, step by step

On the Imaginode canvas, add a Reference node. It contains three fields. First a name: "Léa", "Captain Mist", whatever you like, ideally short. Then a description: that's the text that will be injected into your prompts, and we'll come back to it further down. Finally the photos: drag your best images of the character in there.

This node is reusable everywhere in the project. You create it once, then every Image or Video node on the canvas can plug into it. That's the whole logic of the node canvas: the character becomes a building block sitting somewhere on the board, not a paragraph you copy and paste into twenty different prompts.

One practical detail that changes everyday life: saving is automatic, and every node keeps the history of its last 12 generations. If version 7 of your character was the right one, it's still there, ready to join the Reference node whenever you want it.

The green @mention, or how to summon your character

Once the Reference node exists, you summon the character in any prompt by typing @ followed by its name. The mention shows up in green in the text, a sign that it's properly linked to the node. From then on you write short, natural prompts: "@Léa walks through a night market in the rain, wide shot".

At generation time, Imaginode does two things automatically: it injects the node's description into the prompt, and it attaches the photos as reference images, provided you're using an edit-capable model. Nothing to copy, nothing to re-upload. The mention alone is enough.

A welcome bonus: the magic wand, which rewrites your Image and Video prompts into rich English via Kimi for 1 credit, preserves @mentions. So you can enrich a prompt without breaking the link to your character. It's the kind of detail that spares you nasty surprises at 6 credits per generation.

Which image models can hold a character

On the image side, the champion of references on Imaginode is called Seedream 5 Lite, at 5 credits per generation, roughly €0.05. It's the model to pick by default for any recurring character: it accepts your images and reproduces faces with a fidelity the others don't come close to at this price.

Two alternatives depending on the need. GPT Image Mini, at 2 credits, also accepts references: perfect for iterating on poses and framings without draining the meter. And Flux Kontext, at 6 credits, plays a different role: it edits an existing image. "Replace the background with a beach", and your character stays intact while the scenery changes around them.

On the flip side, Flux Schnell at 1 credit or Imagen 4 Fast at 3 credits are excellent generalists, but don't count on them for the facial consistency of a series hero. The right reflex: rough out the composition with a cheap model, then switch to Seedream 5 Lite with the @mention for the final images.

The master portrait method

Classic mistake: using reference images where the character is already in a scene, small in the frame, backlit, seen from behind. The model then learns as much about the scenery as about the face. The right method is to build a master portrait first: the character alone, on a neutral background, in clean and direct light.

Concretely, open an Image node, pick a photorealistic model like Imagen 4 at 6 credits, and describe only the character: "bust portrait, 30-year-old woman, short red hair, faded denim jacket, plain gray background, soft frontal light". Rerun until you get THE face. Count on 4 or 5 tries, so 24 to 30 credits, less than €0.30.

That portrait becomes the cornerstone of the Reference node. Every image that follows descends from it. It's a ten-minute investment that saves you hours of lottery afterwards, and it's the only step in the whole process where rerunning a lot is genuinely justified.

Vary the angles of your reference photos

A single front-facing portrait is a good start, but the model then knows your character from one angle only. Ask for a profile and it will improvise a nose, a jawline, a pair of ears. To lock the character down in three dimensions, you have to show the model several points of view.

The combination that works well: a front portrait, a three-quarter view, a profile, and a full-body shot for the silhouette and the outfit. Four images are often enough. To produce them, start again from the master portrait with a model that accepts references, and ask for "same person, profile view, same gray background". Check every result before adding it to the node.

Be demanding at this stage, more than anywhere else. A slightly off reference photo, with a subtly different chin or an approximate ear, will silently contaminate every future generation, and you'll go hunting for the cause in the wrong place, in your prompts. Better 4 flawless images than 8 average ones. At 5 credits per view with Seedream 5 Lite, redoing a doubtful photo costs €0.05. Don't negotiate.

Write the description like a continuity contract

The description field of the Reference node is not decorative: it gets injected into every prompt where the @mention appears. So it should contain what must survive every single shot, and nothing else. The durable distinguishing marks: the scar on the left eyebrow, the faded denim jacket, the watch on the right wrist, the hunched walk.

Don't put in anything that changes from one scene to the next. If you write "smiling, in a coffee shop" in the description, your character will be smiling in a coffee shop even in the middle of a storm scene. The description defines identity, the prompt defines the scene. That separation of roles is what makes the whole system robust.

On film sets, this work has a name: the continuity script. Nobody finds it glamorous, and everybody agrees it's what keeps the actor from changing shirts between two shots. Your Reference node plays exactly that role.

And for video: the models and their exact maximums

Video has the same problem as images, only worse: a face that drifts over the seconds is visible immediately. Fortunately, several of Imaginode's video models accept reference images, with different maximums. Seedance 2.0 and 2.0 Fast take 9, Seedance 2.5 even more. MiniMax H3 accepts 9 as well, Grok Imagine Video 5, and Kling V3 and Veo 3.1 Fast take 3 each.

The Video node also exposes separate inputs depending on the model: a start image, and an end image when the model supports it. The most reliable technique remains exactly that one, by the way: generate a perfect image of the character with Seedream 5 Lite first, then plug it in as the video's start image. The motion begins from a face that's already right.

Let's be frank about success rates: AI video fails roughly one generation out of three or four, references or not. A limb passing through an object, a gaze wandering off somewhere. Plan for that waste in your budget. At 26 credits for 5 seconds on Kling 2.5 Turbo, or 48 credits in 720p on Seedance 2.0 Fast, a finished scene costs about double the sticker price in practice. Technical failures, on the other hand, are refunded automatically.

A real case: a six-panel mini comic

Let's take a realistic example: a six-panel mini comic for Instagram, with Louna, a beekeeper, as the heroine. Step 1, the master portrait on a neutral background: 5 tries with Imagen 4, 30 credits. Step 2, the complementary views, three-quarter, profile, full body: about 6 generations with Seedream 5 Lite, 30 credits. The "@Louna" Reference node is ready.

Step 3, the six panels. Each prompt stays short: "@Louna opens a hive at sunrise, high-angle view". With Seedream 5 Lite at 5 credits and counting one retry out of two, the six panels cost about 45 credits. Project total: roughly 105 credits, about €1.05. For a series of six visuals with a stable character, that's hard to beat.

The same node will serve next week for episode 2, then for a 5-second video where Louna waves at the camera. One character, defined once, reused everywhere. That's where the node logic really earns its keep compared with tools where every image starts from a blank page.

Trap number 1: contradictory references

The most common trap: mixing photos from different versions of the character inside the Reference node. Two images where the haircut differs slightly, a third where the jacket is darker. The model, ordered to stay faithful to sources that contradict each other, produces a mushy average that looks like none of the three.

The symptom is easy to recognize: the generations are all "almost right" but never identical to one another. If that happens to you, don't touch the prompt. Open the Reference node, lay the photos side by side, and ruthlessly remove the ones that diverge. Go back to the master portrait as your single source of truth.

Simple rule: every photo you add should be describable by exactly the same sentence as the others. If you have to say "that one is from before she got her scar", it has no business being in the node. Create a second Reference node for that version of the character instead.

Trap number 2: eight nearly identical photos

The opposite trap exists too. Eight front-facing portraits, same expression, same light, same background. The model concludes that everything in them is part of the character's signature: the angle, the frozen smile, the lighting. As a result, it resists when you ask for anything else. Your character stares straight at the camera in every scene, even in a full sprint.

This is overfitting, reference edition: too many identical examples lock in the pose instead of the face. The remedy circles back to the advice on angles: few images, but varied ones. Four photos covering front, three-quarter, profile and silhouette beat eight clones of the same portrait, systematically.

On video models that accept up to 9 references, like Seedance 2.0 or MiniMax H3, the temptation to fill every slot is strong. Resist it. Fill them only if each image brings genuinely new information: an angle, an outfit, an expression. Otherwise, three good images do better than nine redundant ones.

Where to start today

The action plan fits in a single thirty-minute session. Create an Image node, generate your master portrait on a neutral background, budget around thirty credits of tries. Create the Reference node, name the character, write three sentences of durable description, add the portrait and then the complementary views. Then test a first scene with the @mention and Seedream 5 Lite.

If you'd rather not wire all of that yourself, Imaginode's assistant builds the complete workflow in one click for 1 credit per message: describe "a series of posts with a recurring character" and it lays down the Image, Reference and Video nodes already connected. All that's left for you is filling in the photos and adjusting the prompts.

With the Starter plan at €13 excl. VAT for 900 monthly credits, this entire session consumes less than 15% of the plan. And everything happens in the browser, with generation running server-side: nothing to install, no graphics card to own, your lunch-break session on the living room laptop is enough.

Once your character is stable, the next job is staging them properly: framing, lens, camera movement. That's exactly what the Camera node is for, and we've devoted a full article to it, "directing the camera in AI". Your heroine now has a face that holds. Time to give her some real shots.