These are the rules the studio applies on Create video, plus the failure cases that explain why a face changes mid-clip. None of them require a character library — there is not one. You bring the image; the rules decide whether it holds.
Why the reference decides the face
In image-to-video, the model continues the frame it is given. The reference is the only place it can read a face from: the shape of the eyes, the line of the jaw, the distance between them. When the reference is dark, turned away, or too small, the model fills the gaps by inventing — and the invention is a new face.
The motion prompt has no channel for identity. “Keep her face the same” is not an instruction the model can act on; it is a wish. The face in frame one comes from the reference, and every frame after that is the model trying to stay near it.
This is why the order matters. Approving a frame for 6 credits is cheaper than discovering partway through a shot that the face was never going to hold. Frame first, motion second.
Six rules for a reference that holds
A reference that holds has six properties. They all give the model less to invent.
- Front-facing and lit. Eyes visible, light on the face, no hard shadow across it. A turned or shadowed reference forces the model to invent the missing side.
- Half-body or closer. In a full-body wide shot the face is a handful of pixels. Crop to the head and shoulders when the face is what matters.
- One subject. No group photos, collages, screenshots with UI, or two faces. Every extra face is a second identity the model can drift toward.
- Clean background. Plain or softly blurred. Busy patterns and printed text get copied, then crawl as soon as the camera moves.
- Matching aspect ratio. A vertical reference for a vertical shot; a landscape reference fed into a vertical shot gets cropped, often through the face.
- Sharp, not stylised into mush. Soft focus, heavy grain and low resolution hide the eyes. Pick a clearer frame rather than hoping the model guesses.
An uploaded image vs a generated frame
Both image sources work, and they fail differently. A frame you generated on Create was made by the same pipeline as the shot, so its look, grain and resolution already match. An uploaded photo or drawing brings detail the model has never seen, which can help a likeness and is harder to hold.
A shot can use up to three reference images — type @ in the Create video prompt to insert one — and an upload is used for that shot only. Generated frames and finished clips stay in Assets, which is where you find them if a task was still rendering when you closed the page.
Mixing sources has a cost: a reference and a generated frame in different light average together, and the face lands between two versions. Pick one source of truth per sequence.
More references is not more consistency
Three references are not automatically better than one. Three views of the same character from the same shoot can help: a front view, a three-quarter view, a profile. The model gets more angles and fewer excuses to invent.
Three references of different people is the opposite. When they disagree, the model blends them, and the result resembles none. The practical rule: every reference in one shot should be the same person, shot under the same conditions.
If you cannot tell at a glance that two references are the same character, the model cannot either.
Keeping a character across several shots
Consistency between shots is a file habit, not a prompt trick. Keep the reference set identical between takes and change only the motion line. Swapping one reference for a better-lit version changes the face the model is aiming at.
For a sequence, make one frame you are happy with and build every shot from that frame. The character reads as the same person even when the camera changes, because the anchor never changed. Generating a new frame for every shot means generating a new face — that is how reference conditioning works, not a defect.
Lock the look as well as the face: same grain, same palette, same lens language. A perfect face under a different grade still reads as a different show. When an angle needs a new frame, generate it from the same reference image, one character at a time, and check it before you animate it.
What breaks consistency, and the first fix
Five failures that show up most, and the first fix for each.
None of these fixes is a prompt rewrite. The motion prompt can make a shot calmer; it cannot add identity the reference does not contain.
| What you see | What causes it | First fix |
|---|---|---|
| Face changes mid-clip | Reference dark, turned, or small | Replace with a front-lit half-body frame |
| Face resembles nobody | Two references with different people | One subject per shot |
| Face slides toward a generic anime face | Motion is large and fast | Slow the camera; keep subject action small |
| Fine detail boils or crawls | Busy background, fine patterns | Simplify the background; hold the camera still |
| Shots look like different shows | Palette or grain changed between frames | Reuse one frame as the anchor |
The pre-press checklist
Run this list before you press. Each press is a new charge; failed generations are returned automatically, but a take you simply dislike still costs.
The beginner guide covers the rest of the workflow around these rules: frame prompts, motion prompts, and the fix for each common failure.
- One subject, face front-lit and sharp.
- Head and shoulders or half-body, not smaller.
- Plain or blurred background; no text or UI.
- Every reference is the same person, from the same conditions.
- Aspect ratio matches the shot.
- The same frame is reused for every take in a sequence.
Questions
Common questions
- How do I keep a character consistent across several shots?
- Anchor every shot to the same frame and the same reference set, and change only the motion prompt. A new frame is a new face, even when the character description is identical.
- Can I upload my own photo as a reference?
- Yes, if you have the rights to it. It is used for that shot only, and the six rules apply the same way: lit face, half-body or closer, one subject, clean background, matching aspect ratio.
- Why does the face change between takes?
- Takes vary by design. A weak reference or an aggressive camera move makes that variation obvious. Fix the reference first, then slow the motion, before you spend credits on another take.
- Do I need a character library?
- No. FicReel does not have one. You bring a reference image, and the shot is built from that image.
- Can I use a known anime character as a reference?
- No. Do not use third-party characters. Use material you have the rights to; the product does not reproduce someone else’s intellectual property.
- How many reference images can one shot use?
- Up to three, inserted with @ in the motion prompt. More only helps when they are the same character under the same conditions.
Try the workflow on one frame.
New devices start with free credits; nothing spends without your press.
FicReel is an AI anime image and video studio. Prices and wait times on this page come from the same credit table the product uses; if the product disagrees with this page, that is a bug — write to admin@mail.ficreel.com. More: beginner guide, pricing, all guides, glossary, about FicReel. Canonical source: https://ficreel.com/learn/reference-image-guide.