A generated person looks different from clip to clip because every generation draws the face again from whatever it was given. If it was given nothing but a prompt, a different start image each time, or a model that cannot take references, it fills the gaps differently on every run. The fix is the same in almost every case: decide the face once, as a small set of approved stills, and give every clip those same stills, as references on models that take them or as the start frame on models that animate from one image.
TL;DR
- No reference, no continuity. Text-to-video invents the face on every run
- One approved set of stills, not a different start image per clip, and never the last frame of the previous clip
- Reference limits differ by model: Happy Horse and Seedance take up to 9 images on the canvas, Gemini Omni Flash 7, Grok Imagine 10; image models go higher (Nano Banana 10, Seedream 14)
- Kling v3 is image-to-video only on 8frame, so its start frame is the only identity it gets
- Keep one model per character for a sequence, and keep character shots short
- Build a character sheet first: three Nano Banana Pro stills cost 63 credits; four 5-second clips from them start at 196 more
1. No reference attached
With only a prompt, nothing ties clip two to clip one. "A woman in her thirties with short dark hair and a green jacket" describes thousands of faces, and the model picks a new one each run. Adding detail to the prompt narrows the range but does not close it.
Fix: give the model a picture of the character. Generate or choose one clean image of the face first, then attach it to every clip, either as a reference or as the start frame (sections 3 and 4).
2. A different start image per clip
Image-to-video animates the face that is in the first frame. If clip one starts from one portrait and clip two from another generation of "the same" character, you have cast two people. The same thing happens more slowly when you start each clip from a frame of the previous one: whatever changed in clip two is carried into clip three, and so on down the sequence.
Fix: make every start frame from the same approved stills (section 7), and when a shot comes out wrong, go back to those stills rather than to the last clip.
3. The model takes fewer references than you think, or none
Not every video model on 8frame takes reference images, and those that do have different caps:
| Model on 8frame | Images it takes for the character | Notes |
|---|---|---|
| Grok Imagine (reference mode) | 1 to 10 | |
| Happy Horse (reference to video) | 1 to 9 | |
| Seedance 2.5 and 2.0 | up to 9 on the canvas | |
| Gemini Omni Flash (reference to video) | 1 to 7 | |
| Kling O1 Reference | up to 7 | |
| Veo 3.1 multi-reference | up to 3 | 8-second clips only |
| Wan 3.0, Wan 3.0 Prime | 1 start image | Prime's reference mode takes a reference clip, not a set of images |
| Kling v3 Standard and Pro, Kling O3 | 1 start image | no reference images |
Image models, for building the stills: Seedream up to 14, Nano Banana (Classic, 2 and Pro) up to 10, GPT Image up to 10. The full list is in AI reference image limits by model.
One trap: on Seedance, Gemini Omni Flash and Happy Horse you get either a start image or a set of references, not both. If both are attached, the references are sent and the start image is dropped. Seedance's end frame also needs a start image and no reference images.
Fix: for a recurring character, pick a model that takes references, and attach the fewest images that pin the face down: one clean image per angle you need. More references are not automatically better; each one is something else the model has to reconcile.
4. Kling v3 only animates a start frame
On 8frame, Kling v3 (Standard and Pro) is image-to-video only: it needs a start image and takes no reference images. Kling O3 also needs a start image. Whatever face is in that first frame is the only identity the model has, so two clips from two different start frames are two different starting points.
Fix: make each Kling start frame from your character sheet. Put the sheet into Nano Banana Pro as references, ask for the character in the shot's setting and pose, and animate that still. If you want Kling to read several images instead, Kling O1 Reference takes up to 7.
5. Switching models mid-sequence
Each model draws a face its own way, so the same reference can come back as two slightly different people on two models. A sequence that cuts between a Seedance clip and a Kling clip can show that difference at every cut.
Fix: keep one model per character for a sequence. If one shot needs a different model (a longer take, sound, a resolution), regenerate the neighbouring shots on that model too, or put the clips side by side before you commit to the cut.
6. Long clips and big motion
The longer a shot runs, and the more the face turns, moves fast, or goes out of view and comes back, the more chances the model has to drift. We have not measured this per model, so treat it as working advice rather than a rule.
Fix: keep character shots short (5 seconds is the default on most video tools), ask for modest movement and moderate expressions, and cut two short clips from the same references instead of one long take. For the wider problem of things changing mid-shot, see why does my AI video morph.
7. No character sheet
Most of the causes above come back to one missing step: there was never a single, approved version of the face.
Fix: build the sheet before the first clip.
- Start from one reference of the character: a generated portrait or a photo you have the rights to use.
- Generate stills in several angles from it on an image model that takes references: front, three-quarter and profile, plus a waist-up or full-length shot if the clips need the body. On Nano Banana Pro that is 21 credits a still at 1K or 2K.
- Review them side by side and regenerate any still that does not look like the same person.
- Animate only from those stills, as references or as start frames.
To see the character in several moments before spending on video, ShotGrid (beta) makes a 3x3 sheet of nine frames from 1 to 10 reference images for 42 credits.
What a consistent four-clip sequence costs
The same three-still sheet on Nano Banana Pro (3 x 21 = 63 credits), then four 5-second clips:
| Video route | Four clips | Total with the sheet | About |
|---|---|---|---|
| Grok Imagine reference mode, 720p, 3 references | 4 x 49 = 196 | 259 | $2.59 |
| Gemini Omni Flash, reference to video | 4 x 85 = 340 | 403 | $4.03 |
| Happy Horse reference to video, 720p | 4 x 95 = 380 | 443 | $4.43 |
| Kling v3 Standard with sound, from four scene stills (+84) | 4 x 86 = 344 | 491 | $4.91 |
| Seedance 2.5 with references, 720p | 4 x 157 = 628 | 691 | $6.91 |
| Happy Horse reference to video, 1080p (the tool's default) | 4 x 189 = 756 | 819 | $8.19 |
Dollar figures assume a credit at about $0.01, the pack rate. Reference images do not change the price on Seedance, Happy Horse or Gemini Omni Flash; on Grok Imagine each extra image adds a fraction, so a 5-second 720p clip runs from 48 credits with one reference to 50 with ten. Every retake costs the same as the first run, so the sheet is also the cheapest place to catch a wrong face: a still is 21 credits, a clip 49 to 189.
A real person's face
If the character is a real person, you need their permission to use their likeness. The rules for ads are in synthetic talent and celebrity likeness in ads.
FAQ
Why does my AI character look different in every clip? Usually because each clip was generated from a prompt alone, from a different start image, or on a different model. Give every clip the same approved stills, as references or as the start frame.
How many reference images should I use for one character? One clean image per angle the shots need, often three or four. The caps per model are in the table above; staying well under them is fine.
Can Kling v3 use a reference image of my character? Only as the start frame. Kling v3 on 8frame is image-to-video only and takes no reference images; Kling O1 Reference takes up to 7.
Can I use a start image and reference images together? Not on Seedance, Gemini Omni Flash or Happy Horse: with both attached, the references are used and the start image is dropped.
What happens if a generation with my reference fails? 8frame shows "Sensitive content detected. Please try changing your prompt or image." for every failure, whatever the cause, and the credits for a failed generation come back automatically.
Sources
- Reference limits, start-image rules and the start-image-or-references behaviour per model: the 8frame tool configuration and the settings 8frame sends to each provider, read 2026-10-01
- Credit prices (Nano Banana Pro, Grok Imagine, Gemini Omni Flash, Happy Horse, Kling v3, Seedance 2.5, ShotGrid): the 8frame tool registry cost functions, read 2026-10-01
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.