How to Keep a Character Consistent in Pika H3 Flex

The simplest way to keep a character consistent in Pika H3 Flex is to start every shot from the same image and reuse the same description of that character word for word. The model's reference mode offers stronger control, but it needs the local weights or a tool that supports it. Each generation is separate, so expect some drift between clips.

The rules below come from the official prompt guidance on GitHub. The workflow advice is ours; we have not published test clips for it yet, and results vary from run to run.

Which method to use

Using a start frame

1. Make one clear image of the character. Choose a sharp view with the face and outfit fully visible.
2. Upload it as the start frame for every shot in the sequence.
3. Describe what happens next, in order. The guidance says to write actions forward from the opening state, in playback order.
4. Copy one character description into every prompt. Do not reword it between shots.
5. Give each clip one job. Make one action per clip, and change the scene with a new start frame made from the same character.

A 5-second 768p test costs 20 credits, about $0.42–$1.04, so test a look at that length before you commit to longer clips.

A woman in a yellow raincoat with short black hair stands at a bus stop.
She looks up at the rain, then opens a red umbrella. The camera remains fixed.
Sound: steady rain, the umbrella snapping open.

How reference mode works

The open release includes a second adapter, Ref2VA, which takes up to 9 images, 3 videos and 3 audio clips. From the official guidance:

• Label the character once. Define the person as <Subject 1>, name the source image such as <Picture 1>, and use that label everywhere.
• Mark what must be kept. A subject marked fully_preserved keeps its defined features.
• Define everything you care about. Preservation covers only the features you define. If you list face and hair, the outfit is not covered.
• Use a character sheet for looks only. The guidance's example keeps the sheet's panels, labels and alternate views out of the shot.
• Give each reference one role. Do not ask every asset to control everything.

We do not offer reference inputs on this site. See what we do and do not support.

What the guidance does not promise

It has no rules specific to portrait framing or facial detail, and it does not claim identity will stay identical between generations. Its markers express intent; they do not guarantee exact reproduction. For faces, always check the result, and only use a real person's image with their consent.

FAQ

Can Pika H3 Flex keep the same character across videos? It can get close when each shot uses the same start frame and description, or defined references in Ref2VA mode. Each generation is separate, so some drift can remain.

How many reference images does Pika H3 Flex support? The Ref2VA adapter takes up to 9 images, 3 videos and 3 audio clips. This site does not offer reference inputs; it offers start and end frames.

Can I animate a portrait photo with Pika H3 Flex? Yes, upload it as a start frame in image to video and describe the motion. Use photos you have the right to use, and do not imitate a real person without consent.

Should I use text only for a recurring character? Text alone is the weakest option. Use it for quick tests, then move to a start frame.

How do I make a multi-shot sequence? Generate one clip per shot from the same character image, keep the description identical, then join the clips in an editor. See the prompt guide.