Veo 3.1 vs Seedance 2.5: few references or many
Veo 3.1 and Seedance 2.5 can both be called through an API today, and both generate video with audio from reference images. They take opposite approaches. Veo 3.1 accepts three reference images and makes 8-second reference clips. Seedance 2.5 accepts up to 30 images plus video and audio, and makes clips of up to 30 seconds. The right choice depends on how many subjects you need to hold steady, how long your shots are, and whether those subjects are real people.
Status as of October 2026
Veo 3.1 is Google's video model. In the Gemini API, the Veo 3.1 preview model IDs were shut down on October 22, 2026, according to Google's Gemini API docs. Google points developers to Gemini Omni Flash (gemini-omni-1.1-flash) or to the generally available Veo models on its enterprise platform.
Veo 3.1 is also offered through providers such as fal.ai. fal.ai's Veo 3.1 reference-to-video page says that endpoint will be deprecated on November 17, 2026, so check your provider before you build on it.
Seedance 2.5 is a ByteDance model. According to fal.ai's Seedance 2.5 guide, dated August 18, 2026, it is live on fal.ai's API with text-to-video, image-to-video and reference-to-video endpoints.
Reference inputs
Veo 3.1 takes up to three reference images, according to Google's Veo docs as of October 2026. References work on Veo 3.1 and Veo 3.1 Fast; the docs list them as not available on Veo 3.1 Lite. With three slots, you choose the most informative views: usually front, three-quarter and one close-up of the identifying detail.
Seedance 2.5 on fal.ai takes up to 30 images, 10 reference videos and 10 audio clips, 50 files in total, according to fal.ai's guide as of October 2026. Each video or audio file runs 1.8 to 30.2 seconds, and each list is capped at 30.2 seconds combined.
In the Seedance prompt, each file is named as @Image1, @Video1 or @Audio1, so you can say which image is the character, which is the product and which is the location. A reference video of a product turning, or of a character's walk, gives the model motion information that still images cannot.
The practical gap: Veo 3.1's three slots suit one subject per shot. Seedance 2.5 has room for a character, a product and a set in the same request.
Length, resolution and audio
Veo 3.1 generates 4, 6 or 8 seconds, but clips with reference images must be 8 seconds, according to Google's Veo docs as of October 2026. Output is 720p by default, with 1080p and 4K at 8 seconds, in 16:9 or 9:16. Audio is always generated. A clip can be extended in 7-second steps up to 148 seconds in total, at 720p.
Seedance 2.5 generates 4 to 30 seconds in one pass, or a length the model chooses, according to fal.ai's guide as of October 2026. fal.ai lists 480p and 720p output and seven aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 and auto. Audio is generated with the picture, and dialogue written in double quotes comes back synced. Edit and extend run through the same reference endpoint.
If you need 1080p or 4K, Veo 3.1 is the option here. If you need 15 to 30 seconds without a cut, Seedance 2.5 is.
Faces and content rules
We could not find ByteDance's own policy for real faces in Seedance 2.5 references, so we make no claim about it. For Seedance 2.0, provider documentation is explicit. Venice AI's docs as of October 2026 say a detectable human face triggers a consent step, and a recognizable public figure is rejected even with consent.
Google's Veo docs describe reference images as style and content references, subject to Google's safety filters. If your subject is a real presenter or creator, test both models with your references before you commit to either.
Price per clip
On fal.ai, Veo 3.1 reference-to-video costs $0.40 per second with audio at 720p or 1080p, and $0.20 without audio, as of October 2026. At 4K it is $0.60 with audio or $0.40 without. An 8-second 1080p clip with audio is $3.20.
Seedance 2.5 on fal.ai is billed by output tokens, which fal.ai's guide puts at about $0.22 per second at 480p and $0.47 per second at 720p as of October 2026, with audio on or off. An 8-second 720p clip is about $3.80, and a 30-second 720p clip about $14. Image and audio references are not billed; a reference video is.
Which to pick for consistent characters and products
Choose Veo 3.1 for single-subject shots, 8-second beats and 1080p or 4K delivery. Product hero shots, presenter intros and short ad cuts fit it well. Veo 3.1 is the stronger choice when resolution matters.
Choose Seedance 2.5 for scenes with several fixed elements, for motion references and for long single takes up to 30 seconds. It is the more flexible model on inputs and length, if your subject passes its face rules and 720p is enough.
We have not run identical references through both, so this is a comparison of published capabilities and rules, not measured fidelity.
RefVid runs neither model. It runs Kling 3.0 Omni, with 1 to 5 reference photos per subject, up to 3 subjects in one shot and clips of 3 to 15 seconds, and will switch to Kling 4 Omni Reference when its API opens. RefVid is independent and not affiliated with Google or ByteDance.
Frequently asked questions
- Which accepts more reference images, Veo 3.1 or Seedance 2.5?
- Seedance 2.5, by a wide margin: up to 30 images plus 10 videos and 10 audio clips on fal.ai. Veo 3.1 accepts up to three images, per Google's docs as of October 2026.
- Which makes longer clips, Veo 3.1 or Seedance 2.5?
- Seedance 2.5 generates 4 to 30 seconds in one pass. Veo 3.1 clips with reference images are 8 seconds, and can be extended in 7-second steps at 720p.
- Which has higher resolution, Veo 3.1 or Seedance 2.5?
- Veo 3.1, with 1080p and 4K at 8 seconds. As of October 2026, fal.ai lists Seedance 2.5 at 480p and 720p.
- Can I still use the Veo 3.1 preview in the Gemini API?
- No. Google's docs say the Veo 3.1 preview model IDs were shut down on October 22, 2026. Google recommends Gemini Omni Flash or its generally available Veo models instead.