How to keep the same face across different scenes

Character consistency is the thing most image models are worst at, because a text prompt is a poor way to specify a face. The fix is to stop describing the person and start attaching them. OpenModels runs an identity model on our own GPUs that takes your character as a reference image and restages them, and the whole technique is a prompt that describes the destination and never the face.

What actually holds a likeness

A face is a few hundred small relationships — the distance between the eyes, the width of the nose against the mouth, the shape of the hairline — and none of them survive a trip through an English sentence. Every attempt to hold a character with words alone eventually produces a cousin of the person you meant. The reference image is the only thing carrying the identity, so the job of the prompt is to not interfere with it.

Krea 2 Identity is the model built for that on our own hardware. It is a restaging model: it takes the attached character and puts them somewhere else. It requires at least one image and accepts a second, and it has no text-only fallback — the graph will not run without a character to work from, which is a limitation and also the reason it holds a face better than a general model given the same reference.

Which plan: Krea 2 Identity is open on every plan including Free, on our own GPUs, from ⚡1 an image. Its three recipe chips carry no tier gate either. The hosted alternatives — Nano Banana Pro, Seedream 5.0 Pro, the hosted Qwen edits — open from Starter and are paid in ✦ gold.

The three recipes and when each one fits

Krea 2 Identity's built-in chips
RecipeImages you attachWhat it changesWhat it keeps
New outfitOne: the characterClothing you describeFace, pose, location, lighting
New locationOne: the characterSetting, relit to matchFace and outfit
Into a sceneTwo: the scene, then the personPlaces the person in your sceneFace, matched to the scene's light and perspective

The two-image recipe has an order that matters: the scene goes first and the person second. That is the order the model was trained on, and reversing it is the single most common reason a good result turns into a bad one. The composer labels the slots, and dragging them in the tray renumbers the tokens in the prompt with them.

Restaging a character, step by step

  1. Get one clean picture of the character. Front-facing or three-quarter, in focus, with the face large enough in frame to read. This one image is what every later scene inherits, so it is worth generating a few and picking the best rather than using the first.
  2. Attach it and pick Krea 2 Identity. Open /image, drop the picture into the image tray, and tick the model. The recipe chips appear underneath once it is selected.
  3. Choose the recipe and fill in the blank. Each chip loads a prompt with one gap in it. New location, for instance, becomes "create a photo of this person at a rainy night market" — you supply only the destination.
  4. Do not describe the face. No hair colour, no age, no "handsome man with green eyes". Every adjective about the person competes with the reference image and the reference image is the thing that was working.
  5. Run, then reuse the best result as the new reference. Once a scene comes back with the likeness intact, that image is a better reference than your original for anything shot in similar light. Identity drifts slowly if you keep going back to one canonical picture, and fast if you chain result onto result indefinitely.

Open the composer on the identity model — Loads the New location recipe and selects Krea 2 Identity. Attach your character and press Generate yourself.

The multi-reference alternative

The hosted models take a different approach: instead of one identity adapter they accept up to six reference images and reason across all of them. Nano Banana Pro at ✦4.7 an image and Seedream 5.0 Pro at ✦2.2 are the two worth trying, both from the Starter plan. Giving them three or four pictures of the same person from different angles is materially better than giving them one, which is the opposite of how the identity model behaves.

There is a ready-made composer preset for this — "Same face, new scene" — which loads a written instruction and ticks Nano Banana Pro, the hosted Qwen Image Edit and Seedream 5.0 Pro together, so one press gives you three attempts at the same restaging to choose between.

Whichever lane you use, keep one canonical picture of the character in a folder and treat it as the source of truth. Every scene should be generated from that picture rather than from the last scene you happened to like, and when you do promote a result to canonical, do it deliberately rather than by drift.

Where consistency breaks down

If the character is going to be used for months rather than an afternoon, the honest answer is to stop restaging and train a LoRA on them. That bakes the identity into weights instead of re-deriving it from a photo every time, and it removes the reference image from the loop entirely. It needs the Creator plan and it is metered on GPU time, so it is a real commitment rather than a free upgrade.

Can I do this on the free plan?

Yes. Krea 2 Identity runs on our own GPUs, is allowed on every tier including Free, and starts at ⚡1 an image against a wallet that refills to ⚡20 a day. Its recipe chips carry no tier gate.

Why does the model refuse to run with no image attached?

The identity graph restages an attached character and has no text-only path. If you want plain text-to-image on the same checkpoint, that is Krea 2 Turbo, which takes no reference images at all and also ignores negative prompts.

How many reference images should I give it?

One, for the standard restaging recipes. The second slot exists for the two-image case where you supply a scene as well as a person, and the scene goes in first.

Is a LoRA better than reference images?

For a character you will reuse a lot, yes — it holds identity across angles and lighting that references struggle with. It also costs GPU time, needs at least five training photos and is a Creator-plan feature, so it is not the starting point.

Does the same technique work for a product instead of a person?

Yes, and it is the same model. Attach the product, describe the setting, and do not re-describe the product itself.