CHARACTER GENERATION
The same person in every shot
Upload one to three reference photos and generate that character anywhere. New outfit, new pose, new setting, same face. This is the tool that turns a collection of good frames into a sequence that holds together.
The problem this solves
Generate the same person twice from a prompt and you get two different people. Face shape shifts, age drifts, the jawline changes. For a single image that does not matter. For a sequence of shots it is the thing that breaks the work.
Character Generation takes identity out of the prompt and puts it in the reference photos. The prompt then only has to describe what changes, meaning the outfit, the pose, the room they are standing in, while the face stays put.
A fixed seed does not do this. A seed reproduces one image, so as soon as the prompt changes, the face changes with it. Identity has to come from somewhere other than the random starting point.
How character generation works in the form
Two things matter here, and the reference photos matter more than the prompt.
Character Generation sits in the same type dropdown as the other image tools.
Select Character Generation
Character Generation is one of the entries in the generation type dropdown. Choosing it changes the form to ask for reference photos rather than starting from a prompt alone.
Reach for it at the beginning of a project rather than partway through. Establishing the character first is far easier than trying to make a face that already varies across ten frames agree with itself.
Save the character once it works
When a character is behaving the way you want, saving it stores the identity so later generations can call it back rather than starting from the reference photos again.
This is what turns a good result into a usable asset. A saved character is the difference between one strong frame and a sequence that holds together.
Saved actors and characters become reusable across later generations.
Character models
Two models, and the choice is between holding likeness tightly and allowing stylistic movement.
| Model | Best for | Notes |
|---|---|---|
| Seedance 2 Character | Photographic consistency | Holds facial structure across pose and lighting changes. |
| MiniMax Image-01 | Stylistic range | More willing to move the character into different visual treatments. |
Settings reference
What the character form asks for and how each input affects the result.
Reference photos
The images that define the character. Clear, well lit shots of the same person carry identity far better than a mixture of angles and lighting conditions.
1 to 3 images
Prompt
Describes the outfit, pose and setting. Leave facial description out of it, since the reference photos already supply the identity.
Free text
Model
Seedance 2 Character or MiniMax Image-01. The first holds photographic likeness more tightly, the second allows more stylistic movement.
Two options
Choosing reference photos
Three references beat one, but only when they agree. Three shots of the same person in even light from slightly different angles give the model a solid read on facial structure. Three shots taken in different lighting, at different ages, with different hair, give it an argument to resolve.
Heavy shadow across the face is the most common thing that spoils a reference. So is motion blur, and so is a strong colour cast from coloured lighting, since the model reads skin tone from what you give it.
The reference does not have to be a photograph of a real person. A frame you generated with Text to Image works, which is the usual route for an invented character. Design the face first, pick the strongest frame, then use it to lock the identity for everything that follows.
Related tutorials
Longer coverage of character work on the blog.
Character generation questions
Should the prompt describe the character face?
No. The reference photos already carry the identity, and describing the face in the prompt gives the model a second, competing source of truth. Spend the prompt on wardrobe, pose, setting and lighting instead.
Why not just reuse a seed to keep a character consistent?
A seed reproduces one specific image, not one specific person. Change the prompt and the face drifts. Character Generation carries identity across genuinely different prompts, which is what a sequence of shots actually needs.
What makes a good reference photo?
Even lighting, a clearly visible face, and no heavy shadow across the features. Extreme angles, motion blur and strong colour casts all reduce how well the identity transfers.
Can I use a generated image as the reference?
Yes. A frame you produced with Text to Image can serve as the reference photo, which is a common way to design a character first and then lock it in before producing the rest of the shots.
Which of the two models should I use?
Seedance 2 Character when the work is photographic and the likeness has to survive changes in pose and lighting. MiniMax Image-01 when the character needs to move between visual styles and some drift in exact likeness is acceptable.
Other ways to generate images
Character Generation is one of seven generation types in the image workspace.
Video & LipSync
- Video Generator
- Text to Video
- Image to Video
- Start-End Frame to Video
- Draw to Video
- Motion Control
- Video Enhancer
- Video Upscaler
- Video to Video LipSync
- Audio to Video LipSync
- Image to Video LipSync
- Video FaceSwap
- Seedance 2
- Vidu Q3 Pro
- Gemini Omni
- Google Veo 3.1
- Kling 3.0 Pro
- Luma Ray 3.2
- LTX 2.3
- Happy Horse 1.1
- WAN 2.7
- Kling 3.0 Motion
- ByteDance Upscaler
- InfiniteTalk
- InsightFace