A character sheet is a set of views of one character (front, side, back, a few expressions) that later steps use as references. On an 8frame board you make the sheet as stills, either one pose per Nano Banana Pro node or nine frames at once on ShotGrid, then connect those stills to a reference-to-video node such as Happy Horse, which takes 1 to 9 character references, or Grok Imagine, which takes 1 to 10. The limit people do not expect: reference slots and a start frame do not always mix. On Seedance, attaching reference images switches off the end frame, and Wan 3.0's reference-to-video route is not on the canvas at all.
What character consistency is and why models drift is covered in what is character consistency in AI. This page is the board that goes from sheet to clip.
TL;DR
| Step | Node or model on 8frame | Fixed price (credits) | Input it needs |
|---|---|---|---|
| 1. Character concept | Nano Banana Pro or GPT Image 2 | 21 at 1K or 2K; GPT Image 2 high 29 square, 20 other ratios | a description |
| 2a. Sheet, nine views at once | ShotGrid | 42 per 3x3 contact sheet at 4K | 1 to 10 reference images, prompt up to 5,000 characters |
| 2b. Sheet, one view per node | Nano Banana Pro | 21 per image at 1K or 2K, 42 at 4K | the concept still as reference |
| 2c. Sheet, many references | Seedream 5 Lite | 6 per image | up to 14 references |
| 3a. Clip from references | Happy Horse, reference-to-video | 95 for 5 s at 720p, 189 at 1080p | 1 to 9 character references |
| 3b. Clip from references | Grok Imagine, reference mode | 35 at 480p, 48 at 720p, for 5 s | 1 to 10 reference images |
| 3c. Clip from references | Kling O1 Reference | 14.56 per second | references; no end frame |
The board, node by node
Start from CHARACTER POSES, the live template for generating one character in different poses. It covers steps 1 and 2. Add a reference-to-video node at the end for step 3.
Step 1: the concept still
One image that defines the character: face, hair, build, outfit, colour palette. Make it on Nano Banana Pro (21 credits at 1K or 2K) or GPT Image 2 (high quality, the default, 29 credits square or 20 for other ratios).
Keep the character fictional. If the character is based on a real person, that person's written consent is needed for any commercial use, and public figures are off limits.
Full-body concept of a fictional courier character, woman in her twenties, short asymmetric black hair, orange windbreaker over a grey hoodie, black cargo trousers, white sneakers, messenger bag, neutral standing pose, plain light grey background, even studio light, no text
What to fix now: anything you would not want repeated in every clip. Odd hands, a logo on the jacket you do not own, a bag strap that changes sides. Every later image copies this one.
Step 2: the sheet
Two ways, depending on what the video model needs.
Nine views in one image: ShotGrid. ShotGrid makes a 3x3 contact sheet, nine frames in one image, at 4K for 42 credits. It is square (1:1) and takes 1 to 10 reference images and a prompt of up to 5,000 characters. Ask for a turnaround in the prompt: front, three-quarter left, side left, back, side right, three-quarter right, and three close-up expressions. One sheet is one image, so it fills one reference slot.
One view per image: Nano Banana Pro. A node per view, each with the concept still as reference (Nano Banana Pro takes up to 10). Each view is a separate image, so each can fill its own reference slot in the next step. This is the better choice when the video model takes several references, because each slot then carries one clear view rather than nine small ones.
Many inputs: Seedream 5 Lite. Seedream 5 Lite takes up to 14 reference images, the most on 8frame, at 6 credits per image. Use it when the character has many fixed elements (outfit pieces, props, a logo patch) that each need their own reference.
Keep the background, light and scale identical across views. A reference set with mixed lighting tells the video model the character's colours are uncertain.
Step 3: reference-to-video
Connect the sheet images to a reference-to-video node and write the shot.
Happy Horse (Alibaba's model, run through fal) has four modes on 8frame: text-to-video, image-to-video, reference-to-video with 1 to 9 character references, and video edit. Reference-to-video is the one for this board. It is 95 credits for 5 seconds at 720p and 189 at 1080p. The node opens at 5 seconds, 1080p, 16:9. In our production data, a 10 second clip at 1080p had a median generation time of about 195 seconds.
Grok Imagine in reference mode takes 1 to 10 images: 35 credits for 5 seconds at 480p, 48 at 720p. Audio is always on, with no silent mode. It runs 1 to 15 seconds.
Kling O1 Reference bills 14.56 credits per second and takes references, but no end frame.
A shot prompt names the character by what the references show and describes only the action and the setting:
The courier from the references rides a bicycle through a narrow city street at dusk, camera tracking alongside her at handlebar height, orange windbreaker flapping, warm street lights, single continuous shot
Step 4: more shots, same references
Every new shot is a new node connected to the same sheet images. Do not chain clip to clip (taking the last frame of one clip as the reference for the next): small changes add up. Always go back to the sheet.
Join the shots on the Video Editor timeline. The Montage node joins 2 to 10 clips for 10 credits, but its output has no audio track, so Grok Imagine's always-on audio, or Happy Horse's sound, would be dropped.
Which model for which step
| Model | Reference images | Price for 5 s | Resolution | End frame on canvas | Notes |
|---|---|---|---|---|---|
| Happy Horse, reference-to-video | 1 to 9 character references | 95 (720p), 189 (1080p) | 720p, 1080p | not listed | node opens at 1080p 16:9 |
| Grok Imagine, reference | 1 to 10 images | 35 (480p), 48 (720p) | 480p, 720p | no | audio always on; 1 to 15 s |
| Kling O1 Reference | references | 14.56 per second | not listed here | no | also an O1 Image mode with end frame: 73 for 5 s |
| Seedance 2.0 | reference images, or first and last frame, not both | 108 at 720p, 48 at 480p; price ignores references | 480p, 720p | only without references | one reference clip on the canvas |
| Wan 3.0 base, reference-to-video | not on the canvas | API and MCP only | the canvas Wan node sends one start image | ||
| Wan 3.0 Prime | a reference video, on the canvas | 46 / 95 / 189 (480p / 720p / 1080p) | up to 1080p | API only | a video of a performance, not a sheet of stills |
For stills: Nano Banana Pro and Nano Banana 2 take up to 10 references, GPT Image 2 takes 1 to 10, Seedream 5 Lite up to 14, Grok Image 2.0 up to 3.
Where it breaks
- Too many tiny views in one slot. A 3x3 sheet in one reference slot gives the video model nine small faces. If the model takes several references, feed one view per slot.
- Drift by chaining. Taking the last frame of clip one as the reference for clip two copies its errors. Reference the sheet every time.
- Mixed lighting across views. Re-make the odd view, not the clip.
- Expecting an end frame with references. On Seedance it is references or first and last frame, not both. Kling O1 Reference takes no end frame.
- Wan reference-to-video. Not on the canvas. The canvas Wan node sends one start image; only Wan 3.0 Prime takes a reference video there.
- Unwanted audio. Grok Imagine always adds audio. Mute it on the Video Editor timeline.
- The generic error. Most failures show "Sensitive content detected. Please try changing your prompt or image." Failed runs are refunded automatically. For Seedance specifically: Seedance 2.0 and 2.5 refuse reference images with a large photoreal close-up face (error E005); waist-up and full-body references pass; retrying does not help.
What other tools call this
Character sheets are a common board pattern. Flora's techniques gallery lists "Character Lock", "Character Variator" and "Character creator"; it publishes no usage numbers. The difference on 8frame is that the sheet and the video model sit on the same board, and each model's reference limit is fixed: 9 on Happy Horse, 10 on Grok Imagine, 10 on Nano Banana, 14 on Seedream 5 Lite.
FAQ
What is a character sheet in AI video? A set of stills of one character from several angles, with a few expressions, used as references so the video model keeps the character the same across shots.
How many reference images can I use? Happy Horse takes 1 to 9 character references, Grok Imagine 1 to 10. For stills, Nano Banana 10 and Seedream 5 Lite 14.
Should I use ShotGrid or separate stills? ShotGrid makes nine views in one 4K image for 42 credits, which is quick and fills one reference slot. Separate stills cost more runs but give each reference slot one clear view.
Can I use a start frame and references together? It depends on the model. On Seedance, attaching references turns off the end frame. Check each node's inputs.
Does Wan 3.0 do reference-to-video on 8frame? Not from stills on the canvas. The canvas Wan node takes one start image; Wan 3.0 Prime takes a reference video.
Sources
- 8frame prices, defaults and limits: the 8frame app, checked 2026-11-09.
- Flora techniques gallery: https://flora.ai/techniques, checked 2026-11-09.
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.