Character consistency is the hardest problem in AI video — and reference images are the solution. A good reference workflow keeps your protagonist recognizable across every shot of a project. This guide covers the complete method: creating references, using them effectively, and fixing drift.
Why references work
Video models generate each clip independently — they have no memory of your previous shots. A reference image gives the model a visual anchor: “this face, this jacket, this look.” Combined with identical text descriptions, references are the closest thing to casting the same actor in every scene.
Step 1: Create your character reference
- One clear image. Front-facing or three-quarter view, neutral expression, good lighting, plain background. The model needs to see the face clearly — dramatic angles and shadows make bad references.
- Show the full look. If the outfit matters, a second full-body reference helps. Keep it to 1–3 images; more references can confuse some tools.
- Generate or photograph? Either works. Many creators generate the reference with an image model first (iterating until the design is right), then use it for all video. A real photo works too if you have the rights.
- Lock it. Once chosen, this image is canon. Don’t “improve” it mid-project — every change ripples into inconsistency.
Step 2: Write the canonical description
Alongside the image, write one fixed text description and use it verbatim in every prompt:
“Mara: woman in her early 30s, sharp jawline, dark brown eyes, black bob haircut, faded denim jacket over a gray t-shirt.”
Image + identical text is stronger than either alone. The image anchors the visuals; the text anchors the model’s understanding. Copy-paste — never retype, never paraphrase.
Step 3: Use references per shot
- Attach the reference to every generation involving the character, in image-to-video or character-reference modes (Kling, Hailuo, and Seedance all support this; check your tool’s exact mechanism).
- Keep prompts otherwise consistent: same style words, same lighting family, same template structure.
- One character per reference set. Multi-character scenes: attach each character’s reference and name them distinctly in the prompt.
- Match the reference’s logic: don’t ask a daytime reference to play a nightclub scene without acknowledging the lighting shift — describe how the look adapts (“same face and jacket, now under neon club lighting”).
Step 4: Check and correct drift
Even with references, drift happens. Check every generation against the reference:
- Face check: compare eyes, face shape, and hairline. Small changes compound across shots.
- Wardrobe check: jacket color, patterns, accessories — the details viewers track subconsciously.
- When drift appears: regenerate with stronger reference weighting (if your tool offers it), simplify the shot, or move the character slightly further from camera — distance hides drift.
- The 3-shot rule: if a character survives 3 consecutive shots recognizably, your system works. If not, fix the reference or description before generating more.
Advanced: reference discipline for series
- Reference bible: one document with every character’s image + canonical description + wardrobe notes. Anyone (or any future you) can generate consistently from it.
- Version control: if a character’s look evolves, save the new reference as v2 with a date. Never mix versions in one project.
- Location references too: the same method works for recurring locations — one anchor image + fixed description keeps your world coherent.
- Test new tools against the bible: when trying a new generator, run your existing references through it first. If it can’t hold your character, it’s not your tool.
The honest limits
References dramatically improve consistency; they don’t guarantee it. Extreme angles, heavy motion, and crowded scenes still cause drift. Plan around the limits: favor medium shots over extreme close-ups for critical character moments, keep action readable rather than chaotic, and budget regeneration attempts for the hardest shots. Consistency is a system — references, descriptions, shot discipline, and editing — not a single feature.
Part of the consistency toolkit alongside the character consistency guide and multi-shot storytelling.