Veo 3.1 reference images and extension: Google yes, Sume text-only

On Google's Gemini API, Veo 3.1 accepts up to 3 reference images and can extend a Veo clip by 7 seconds, up to 20 times (read 2026-10-10). Sume's Veo rows do neither: in the API code they are text-to-video only and refuse frame, reference and video inputs.

If your Veo plan depends on references or extension, Sume's Veo rows will not do the job. Use a different Sume row for references, and keep extension work on Google.

What Google's page says

The Gemini API Veo page (read 2026-10-10) lists three ids: veo-3.1-generate-preview, veo-3.1-fast-generate-preview and veo-3.1-lite-generate-preview. Extension works on 3.1 and Fast only, adds 7 seconds each time, and can repeat up to 20 times. Reference images cap at 3.

Google stores generated videos for 2 days on that page, so download or extend inside the window.

How Sume's Veo rows behave

Sume's request builder for Veo (buildVeoRequest in the API code) takes the prompt, a duration of 4, 6 or 8 seconds, 720p, an aspect ratio of 16:9 or 9:16, and an optional audio flag. It throws on frame fields, reference fields and video fields. The catalog description reads "text-to-video only, 720p, 4/6/8 seconds, 16:9 or 9:16, with optional synchronized audio."

These rows are switched on per environment. They show up in your workspace only if GET /v1/videos/models returns the id, so check the catalog and do not assume either way. An earlier post, linked below, covers the same availability question.

What to use for references on Sume

The Video Generation docs say Seedance 2.x, Wan 3.0 and the MiniMax H3 rows accept image, video and audio references, and Gemini Omni Flash 1.1 accepts image and video references. In the schema, Wan 3.0 takes up to 10 images, 5 videos and 5 audio files. Omni takes up to 10 images and 3 reference videos of at most 3 seconds each.

None of these is Veo, so look at your reference clip first. If you need Google's look, stay on Google. If you need consistency with a character sheet, a reference-capable row is the practical stand-in.

• References wanted: Wan 3.0, Seedance 2.5 or Omni.
• Extension wanted: Seedance 2.5 or Wan 3.0 reach 30 s in one request.
• Veo look wanted: use Google's API.

A longer shot without extension

Extension at Google adds 7 seconds per step. On Sume you get length in one request instead: Wan 3.0 and Seedance 2.5 both accept up to 30 seconds. The table compares an 8-second Veo clip with a 30-second shot on those rows, using Sume's billed rates (list times 1.25) and the Veo row only if it is listed in your workspace.

Cost is not the only difference. A single 30-second request has no seams, while a chain of 7-second extensions inherits the previous clip's last frames. But you also lose Veo's look, so treat this as a substitute, not a swap.

Checking support in the catalog

Reference-heavy work is where the choice matters most. Google caps references at 3 images and gives you extension; Sume's reference-capable rows allow more inputs but no extension step. If the 3-image limit was your constraint, Wan 3.0's 10 images and Omni's 10 images are a real gain, and Seedance, Wan and MiniMax also take audio references that Veo's page does not list.

Whatever row you choose, read supported_input_references and supported_frame_images from GET /v1/videos/models before you send reference files. A model accepts a reference type only if the list includes it, per the Video Generation docs.