Reference-Guided Video Generation with Character Consistency
Generate AI video from character reference images so the same person keeps their face, outfit, and look across shots — no face-swapping for serial IP content, with multiple shots stitched into one continuous clip.

Tool names in this guide are for reference only and may have changed or been renamed. Please use the latest tools in the tool marketplace when implementing.
Key takeaways
- Lock a character with reference images so face and outfit stay on-model
- No face-swapping across serial IP shorts — one character across many clips
- Stabilize a single shot, then stitch multiple shots into a continuous clip
Tools used
Why characters keep "swapping faces" in video
For serial IP shorts and continuous story content, the hard part isn't generating a clip — it's keeping the same character consistent across different shots. Pure text-to-video reimagines the person every time: a rounder face in this shot, a new hairstyle in the next, drifting outfit colours. Stitch a few shots together and it looks like three different actors. For anyone building a continuous narrative, that "face swap" destroys the character's recognizability and the viewer's immersion.
Reference-guided video flips the approach: you give the model a few character reference images and have it generate video anchored on them, instead of imagining the person from text alone. An AI agent connected to reference-to-video tools turns this into a single-sentence operation — upload the references, describe the action, and keep the character locked.
The workflow in four steps
- Prepare character references: pick a few clear reference images of the same person.
- Reference-to-video: generate a clip anchored on the reference images.
- Verify character consistency: compare face and outfit against the references.
- Stitch multiple shots: once stable, generate several shots and stitch them into a continuous clip.
How to do each step
Step 1: Prepare character reference images
Give the agent a few reference images of the same person — front-facing, clean lighting works best. Angles and expressions can vary, but keep it the same person in the same look so the model can extract stable character features.
use ae to generate a video of this character walking, keep the face and outfit consistent with the reference images
Step 2: Reference-to-video
The agent calls a reference-to-video tool to generate a clip anchored on the references. For multi-image character consistency, use ae_pixverse_c1_reference_to_video; when you need more general video generation, use ae_seedance_2_video_generation alongside a reference description. Spelling out "stay consistent with the reference" in the instruction matters.
use ae to render the clip with ae_pixverse_c1_reference_to_video
Step 3: Verify character consistency
Compare the output against the references item by item: face shape, features, hairstyle, outfit colours. This is the main difference from pure text-to-video — reference-to-video has an anchor to align to, while text-to-video imagines anew each time. If it drifts, add a clearer reference or emphasize the details to lock in your instruction.
Step 4: Stitch multiple shots
Once a single shot is stable, use the same reference set to generate several shots — different actions, different framings, but always the same character. Confirm consistency across shots, then stitch them in narrative order into one continuous clip.
Text-to-video vs reference-to-video:
| Pure text-to-video | Reference-to-video (character-consistent) |
|---|---|
| Reimagines the person per shot, prone to face swaps | Anchors on reference images, locks the character |
| Outfit / hairstyle drift between shots | Can require face and outfit to stay consistent across shots |
| Multiple shots look like different actors | One character runs through every shot |
Tools by goal:
| Goal | Tool (example) |
|---|---|
| Multi-image character consistency | ae_pixverse_c1_reference_to_video |
| General video generation | ae_seedance_2_video_generation |
How to get started
- Sign up — new users get 1,000 free credits.
- Connect in-app from an agent like WorkBuddy or Trae by searching for "AgentEarth", or paste a one-line MCP config into Claude Code, Cursor or VSCode.
- Upload character reference images, describe the action in natural language, and let the agent do reference-to-video.
FAQ
How many reference images do I need? One can work, but a few references of the same person from different angles help the model extract stable features and improve cross-shot consistency.
Can it keep the outfit and face? Yes. State in your instruction that face, hairstyle, and outfit colours must stay locked, then verify against the references after generation; add a clearer reference if it isn't tight enough.
Does it support multiple shots? Yes. Stabilize a single shot first, then use the same reference set to generate several shots with different actions and framings, and stitch them in narrative order.
