How to Choose Reference Images for Consistent AI Video Characters
Choose better AI video reference images: prepare a clear character photo, check all four views, separate outfits and settings, and fix drift before filming.
For consistent AI video characters, choose a reference image that clearly shows one subject, then check the generated character from every angle before making video. In Paintbrush, a photo or drawing guides a reusable character with a front view, side profile, back view and face close-up. Mention that character in each shot to use its references again. A clear reference cannot guarantee a perfect take, but it gives you something specific to inspect and correct before an expensive sequence depends on it.
This guide covers choosing the source image, checking the generated views, separating a character from their clothes and location, and diagnosing a reference problem. The goal is a small set of useful visual decisions that hold together across a conversation, an entrance and a reverse angle.
What makes a good character reference image?
Start with an image where the subject is easy to identify. Prefer a clear face, visible hair, readable clothing and enough of the body to understand the silhouette. An unobstructed, evenly lit photo is easier to interpret than a dark party picture or a distant figure in a landscape. For a drawing, choose a finished design whose colors and proportions you actually want to keep.
You do not need a perfect studio photograph. You do need to notice what the image leaves unspecified. A headshot says little about trousers or shoes. A person behind a table has an uncertain lower-body silhouette. Describe those missing details when you create the character so the first generated front view has a clear target.
- Choose one subject. Crop a group photo to the person you want when possible. If other figures remain, identify the intended subject clearly in the character description.
- Keep identity details visible. Glasses, hair shape, facial markings and a distinctive jacket are useful anchors. Heavy shadows and obstructions make those details harder to read.
- Avoid contradictory examples. Different ages, hairstyles or costumes create decisions you may not want the model to make for you.
- Use a reference you can evaluate. If the source face is tiny, you will struggle to judge whether the generated close-up resembles it.
For the upload workflow and a complete example, follow turning a photo into an animated character. The reference can also be a drawing; the important part is a recognizable design.
What should you write alongside the reference?
Describe the character you want, including details the source does not show. Keep action and location out of this description. The character asset should depict one person or creature standing alone; the shot description is where you later ask them to run, sit, argue or open a door.
Mara, the woman on the left in the reference. Keep her short dark curls, round glasses and freckles. Full-body character wearing a mustard cardigan, dark blue trousers and brown ankle boots. No other people.
Choose the project style deliberately. A photograph supplies identity information; it does not require a photoreal video. When transforming someone into a cartoon, judge whether the recognizable features survived in that style, rather than expecting the rendering to reproduce every photographic detail. A clear silhouette and a few distinctive features are more useful than a long list of tiny accessories.
How do you check the four generated views?
Paintbrush generates the front view first, then derives the side, back and face views from it. Check the front before deciding that the character is ready, and inspect all the other views as well. They should describe the same person in the same clothes. A good front view does not excuse a side view with different hair or a back view with a different coat.
| View | What to check | Why it matters |
|---|---|---|
| Front | Face, silhouette, clothes, shoes and one subject only | It establishes the design used to create the other views |
| Side | Nose, jaw, hairstyle, proportions and outfit | Conversations often show someone in profile |
| Back | Hair length, jacket shape, colors and accessories | Entrances and over-the-shoulder shots expose the rear design |
| Face | Eyes, glasses, freckles and facial proportions | Close-ups make identity differences particularly noticeable |
Write a concrete correction when something is wrong: "keep the round glasses and shorten the hair to the jaw" is easier to evaluate than "make her more accurate." Use Keep current look for a small change to a design you already like. If the current rendering style itself is the problem, keeping that look can preserve the problem; make the intended style and source reference clear instead.
Should a new outfit be a new character?
Use an outfit for the same character in different clothes. In Paintbrush, an outfit gets its own front, side, back and face images, and you choose it for the chapter. That gives every shot in the chapter the same dressed version of the character. The original look remains available for other chapters.
For example, Mara can wear her cardigan at the cafe and a raincoat outside. Create the raincoat outfit and check its views before generating the exterior chapter. Asking one shot to "put Mara in a raincoat" while supplying her cardigan references creates conflicting instructions. Our outfit guide walks through the chapter selection.
How should you prepare the setting?
Keep the location separate from the cast. Create a setting that shows the empty place, with the layout, materials and lighting that matter to the story. A character reference should not have to establish the cafe's furniture, and the cafe reference should not contain a second copy of your lead.
Think about what each camera angle will reveal. A wide view of the front counter may not establish the booth beside the window. If that area matters, describe it clearly and consider a separate setting for that part of the place. Reusing a room reference helps, but it cannot specify every wall outside its frame. See keeping backgrounds consistent for examples.
How do references reach the video shot?
Use the character and setting @mentions in What it shows. Select the intended assets from the suggestions rather than relying on a name typed as ordinary prose. Describe the camera, visible people, positions and action. Put spoken lines in the chapter script, with their speakers.
Medium shot inside @Corner Cafe. @Mara sits beside the window, holding a closed notebook. @Theo stands across the table. Mara looks up as Theo pulls out the chair.
Only include characters and places that the shot needs. A reaction close-up does not need the entire cast visible. This makes the composition easier to judge and gives each character a clear role. More references are not a substitute for a specific shot description, and a crowded image can still blend similar faces or clothing.
Is the problem the reference or the shot?
If the same wrong haircut appears in several shots, inspect the character views first. If the character looks correct elsewhere but is misplaced in one scene, inspect that shot's staging. If the clothes change at a chapter boundary, check the selected outfit. These are different fixes.
After changing an asset, existing videos still show the old generated design. Regenerate the affected shots deliberately and review the sequence in order. Shots staged from an earlier ending can inherit that ending's old clothes, posture or prop placement. Fix the earliest affected shot before building later takes on it. The shot repair guide explains when to regenerate, tweak or trim.
What does preparing references cost?
As of October 2026, Paintbrush's base price is 22 credits for a four-view character or outfit and 4.5 credits for a setting. These images use Seedream 5.0 Pro. Asset edits with more than one reference add 0.3 credits for each reference after the first. Two characters and one setting therefore start at 48.5 credits, before extra outfits, revisions or speech.
For comparison, a 5-second video shot costs 151.5 credits with on-camera dialogue or 164 credits when silent or narrated, before a keyframe. Checking the images first can prevent spending video credits on a design you already know is wrong. Use the generation estimate and the full cost breakdown to budget the sequence.
Frequently asked questions
Do I need to upload four photographs?
No. Paintbrush generates the additional character views from the front view it creates. Start with a useful source photo or drawing and inspect the resulting views rather than assembling a collage of unrelated poses.
Will the same references guarantee an identical face in every frame?
No. References guide generation, but motion, occlusion, camera angles and multiple people can still produce drift. Review each take and correct the earliest visible problem before generating dependent shots.
Can I reuse the character in later shots?
Yes. Mention the existing character again in the project. You do not need to recreate its reference sheet for every shot. Use chapter outfits when the clothes change, and preserve the approved character design when they do not.
Build your cast before your first shot.
Create characters, plan scenes, and bring your story to life in one workspace.
Try Paintbrush