Reference-Based Generation
Ciaro keeps faces, props, and places consistent with reference images — not by training a custom character or style model.
Reference-based generation
Ciaro keeps faces, props, and places consistent with reference images — not by training a custom character or style model.
You prepare pictures in Entity Builder (or Asset Manager), drop them into Create View slots, and mention them in the prompt with / chips. The image model copies what it can see in those pictures while Art direction handles the project look.
The loop
- Build a portfolio for the character, prop, or location (Create Image, upload, or Image editor). Pick a clear selected image.
- Open the shot in Create View.
- Confirm Art direction (the look of the board).
- Drag portfolio images from Asset Manager into numbered reference slots.
- In the prompt, type / and insert chips for those people, objects, and slots.
- Describe this shot (action, camera, this moment’s lighting) and generate.
That is the whole consistency system. There is no separate “train my character” or “train my style” step.
What belongs in a slot
| Slot is for | Typical source |
|---|---|
| Who | Character portrait from Entity Builder |
| What object | Prop portfolio image |
| Where | Background / location image |
| Framing help | A sketch, a previous shot, or a composition still |
Use a few strong pictures. Mixed, cluttered, or contradictory references make the model invent a new face or object.
You usually do not put a style image in a slot. Art direction already applies the look to shot images. A style slot is only for a deliberate extra treatment on this one generation.
Prompt chips
Type / in the prompt editor to insert a character, prop, or slot chip. Chips beat renaming people in prose (“the woman in the red coat”) because they stay tied to the prepared asset.
Leave the drawing medium to Art direction. References carry identity; Art direction carries how the board is drawn.
Why this works better than training
- You can change a face or outfit by swapping the picture — no wait for a training job.
- The same hero can appear in cinematic or illustrated boards because portraits stay identity photos; the shot generator applies the look around them.
- Studio crews reuse the same portfolio across episodes from the workspace library (see Asset Manager).
When faces or objects drift
- Is a clear portrait actually in a slot (not only named in the prompt)?
- Is the selected Entity Builder image the right person or object?
- Are you mixing several different faces for one character?
- Did you restack a new medium in the prompt on top of Art direction?
Fix the pictures and slots first. See Character Image Consistency, Using characters in shot prompts, and Art Direction.
Video
Video starts from the approved still. Motion prompts describe movement and camera. Do not try to “re-train” a look in the motion prompt — the start frame already holds identity and Art direction.