Correction, 2026-10-01: this page said Seedance 2.5 also takes up to 10 reference clips "of up to 30 seconds each" and up to 10 audio tracks. On the 8frame canvas, Seedance takes one reference video and one audio track. The 10 clips and 10 tracks are the model's own limits at its provider, and there the clips are capped at 30 seconds combined, not 30 seconds each. The video table also left out Veo 3.1 multi-reference (up to 3) and the video models that take no reference images at all (Wan 3.0, Kling v3, Kling O3), and the page did not say that Seedance, Gemini Omni Flash and Happy Horse take a start image or references, not both. All of this is added below.
Among the models on 8frame that take reference images, the cap ranges from 3 to 14. For images, Seedream takes the most (14), then Nano Banana and GPT Image (10 each). For video, Grok Imagine's reference mode takes 10, Happy Horse 9, Gemini Omni Flash and Kling O1 7, Veo 3.1 multi-reference 3. Seedance is the odd one: version 2.5 accepts up to 30 at its provider, but the canvas lets you attach 9. Some video models take none: Wan 3.0 and Kling v3 animate one start image.
TL;DR
- Image models: Seedream 14, Nano Banana 10, GPT Image 10, Grok Imagine Image 3, Qwen Image 3 (edits) 3.
- Video models: Grok Imagine reference mode 10, Seedance 9 on the canvas, Happy Horse 9 (5 when editing a video), Gemini Omni Flash 7, Kling O1 Reference 7, Veo 3.1 multi-reference 3.
- No reference images: Wan 3.0, Kling v3 (Standard and Pro) and Kling O3 take one start image.
- Start image or references, not both, on Seedance, Gemini Omni Flash and Happy Horse. With both attached, the start image is dropped.
- Seedance on the canvas also takes one reference video and one audio track. The model accepts more at its provider.
- More references is not automatically better. Each one is another thing the model has to reconcile; use the fewest that pin down what must not change.
Image models
| Model | Max reference images |
|---|---|
| Seedream (5 Pro, 5, 4) | 14 |
| Nano Banana (Classic, 2, Pro) | 10 |
| GPT Image | 10 |
| Reve (switched off since 2026-10-07) | n/a |
| Grok Imagine Image | 3 |
| Qwen Image 3 (editing) | 1 to 3 |
Grok Imagine Image's 3 is the provider's own cap, not an 8frame setting.
Video models
| Model | Reference images on the canvas | Other inputs on the canvas |
|---|---|---|
| Grok Imagine (reference mode) | 1 to 10 | |
| Seedance 2.5 | up to 9 | 1 reference video, 1 audio track |
| Seedance 2.0 | up to 9 | 1 reference video, 1 audio track |
| Happy Horse (reference to video) | 1 to 9 | |
| Happy Horse (video edit) | up to 5 | the video being edited, 3 to 60 s |
| Gemini Omni Flash | 1 to 7 | |
| Kling O1 Reference | up to 7 | |
| Veo 3.1 multi-reference | up to 3 | 8-second clips only |
| Wan 3.0 | none | 1 start image; Wan 3.0 Prime's reference mode takes a reference clip |
| Kling v3 Standard and Pro, Kling O3 | none | 1 start image, required |
At its provider, Seedance 2.5 accepts up to 30 reference images, up to 10 reference videos and up to 10 audio tracks, with the videos capped at 30 seconds combined and the audio at 30 seconds combined. Seedance 2.0 accepts up to 9 images, 3 videos and 3 audio tracks, each capped at 15 seconds combined. On the 8frame canvas you get the 9 images, one video and one audio track.
Happy Horse's reference-to-video takes 1 to 9 character references. It costs 95 credits for 5 seconds at 720p and 189 at 1080p.
Start image or references
On Seedance, Gemini Omni Flash and Happy Horse, a start image and reference images do not combine. 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.
Kling v3 (Standard and Pro) is image-to-video only on 8frame: it needs a start image and takes no reference images. Kling O3 also needs a start image. With these, the start frame is the only identity the model gets, so if a character must match across clips, make every start frame from the same approved stills.
Which references to use, and how many
References do three different jobs, and it helps to decide which job each image is doing before you attach it:
- Identity: the person, character or product that must look the same. Start with one clean image per subject.
- Style: a palette, lighting or finish to copy. One or two frames.
- Layout: a composition or a start frame. Usually one.
A practical ceiling for most shots is one image per subject plus one for style. Add more only when a specific detail keeps drifting, and then add the image that shows that detail.
Faces as references
If a reference shows a face in close-up, Seedance's filter may reject the run; see Seedance "flagged as sensitive" (E005). The other size and format rules for reference images are in upload limits by model.
FAQ
How many reference images can Nano Banana use? Up to 10.
How many reference images does Seedance accept? On the 8frame canvas, 9 for both 2.0 and 2.5. At its provider, Seedance 2.5 accepts up to 30 and 2.0 up to 9.
Which AI video model takes the most reference images? Grok Imagine's reference mode, with 10. Happy Horse and Seedance take up to 9 on the canvas.
How many reference images does Gemini Omni Flash take? Between 1 and 7.
Does Kling v3 take reference images? No. On 8frame, Kling v3 is image-to-video only and takes one start image. Kling O1 Reference takes up to 7 reference images.
Can I attach a start image and reference images together? Not on Seedance, Gemini Omni Flash or Happy Horse: the references are used and the start image is dropped.
Sources
- Reference caps on the canvas, start-image rules, the start-image-or-references behaviour of Seedance, Gemini Omni Flash and Happy Horse, and Seedance's reference video and audio inputs: the 8frame app, checked 2026-10-01
- Happy Horse prices: the 8frame app, checked 2026-10-01
- Seedance 2.5 limits at the provider (30 images, 10 videos and 10 audio tracks, 30 seconds combined): replicate.com/bytedance/seedance-2.5/api/schema, checked 2026-10-01
- Seedance 2.0 limits at the provider (9 images, 3 videos and 3 audio tracks, 15 seconds combined): replicate.com/bytedance/seedance-2.0/api/schema, checked 2026-10-01
- Seedance close-up face refusals (E005): 8frame production runs, 2026
One reference per subject, one for style, and add more only when a detail drifts.
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.