Reference limits in 2026 video models: Seedance, Wan, Omni, H3
Seedance 2.5 takes 30 images, 10 clips, 10 audio files per pass; Wan 3.0 up to 20 assets; Omni video refs of 3 s. H3 numbers and Sume's rows.

Vendor pages for 2026 video models give very different reference limits: Seedance 2.5 lists up to 30 images, 10 video clips and 10 audio clips per pass, Wan 3.0 lists up to 20 reference assets, and Gemini Omni Flash 1.1 lists video references of up to 3 seconds. A third-party summary of MiniMax H3 lists 9 images, 3 videos and 3 audio clips.
What you can send through Sume is set by the catalog row, so read supported_input_references for the id before you design a brief around a vendor number.
What each vendor says
The ByteDance Seed page for Seedance 2.5 (read 2026-10-05) says a pass takes up to 30 images, 10 video clips and 10 audio clips, with up to 30 seconds per generation. The Wan 3.0 GitHub page (read 2026-10-05) lists up to 20 reference assets, including documents, webpages, text and images, native 30-second output and up to 12 sequential images with a cohesive style.
Google's Omni 1.1 Flash post (read 2026-10-05) lists video references of up to 3 seconds. The earlier Omni Flash release, per Google Cloud (read 2026-10-05), took up to 7 image references and up to 3 clips of 3 seconds or less.
MiniMax H3
MiniMax's own announcement (read 2026-10-05) describes reference across image, video and audio without stating counts. The counts of 9 images, 3 video clips and 3 audio clips come from a third-party summary on Hugging Face (read 2026-10-05), not from MiniMax, so treat them as a pointer. On Sume, the Video Router notes add one pricing fact: minimax-h3 charges $0.08 list per image after the fifth, $0.10 billed.
| Model | Images | Video clips | Audio clips | Source |
|---|---|---|---|---|
| Seedance 2.5 | 30 | 10 | 10 | ByteDance Seed |
| Wan 3.0 | 20 assets in total (mixed types) | Not itemized | Not itemized | Alibaba GitHub |
| Gemini Omni Flash 1.1 | Not stated on the page read | Video refs up to 3 s | No audio ref on Sume | |
| Gemini Omni Flash (first release) | 7 | 3 clips of 3 s or less | Not stated | Google Cloud |
| MiniMax H3 | 9 | 3 | 3 | Third-party summary |
What Sume's catalog says
The Video Router docs describe the Omni 1.1 row as reference-to-video with up to 10 images and up to 3 clips of 3 seconds or less, addressed as <IMAGE_REF_0> and <VIDEO_REF_0>, and with no reference_audio_urls. Genjutsu takes 1 to 8 reference images and Recast 1 to 4 person photos. The kling-3 row has no reference_*_urls at all.
The docs for H3 and H3 Max say they accept image, video and audio references, and that the Seedance 2.x models and Wan 3.0 do too. They do not restate the vendor counts in the table, so a vendor limit is a ceiling, not a promise that Sume passes it through.
- Design the brief for the smallest limit you must support.
- Read
supported_input_referencesfor the exact id. - Ask for a cost quote before a large reference brief on
minimax-h3, which adds $0.10 per image after the fifth.
Designing a brief around the smallest limit
If you plan to run the same brief on more than one model, count references to the smallest limit in play. A brief with 3 images and one 3-second video clip fits Omni 1.1 Flash, Wan 3.0, Seedance 2.5 and H3 alike, and avoids the extra per-image charge on minimax-h3. A brief with 12 images fits only the wider models, and you should check the row on Sume before you build it.
Reference audio is the narrowest capability. The Sume docs say Gemini Omni Flash 1.1 accepts image and video references but not audio, while H3, H3 Max, Wan 3.0 and the Seedance 2.x rows accept audio references. If your brief has a voice or music reference, that already removes Omni from the shortlist.
Sources
- ByteDance Seed: Introducing Seedance 2.5 (read 2026-10-05)
- Alibaba Cloud: Wan 3.0 on GitHub (read 2026-10-05)
- Google: Build with Gemini Omni 1.1 Flash (read 2026-10-05)
- Google Cloud: Gemini Omni Flash availability (read 2026-10-05)
- Hugging Face community blog (third party): MiniMax H3 summary (read 2026-10-05)
- Video Router API
Related posts
More in Comparisons
- Vidu Q2 image prices: $0.03 text-to-image vs Sume image rates
Vidu lists Q2 text-to-image at $0.03 per 1080P image and reference-to-image at $0.04. Sume image models run $0.025 to $0.275. Where they overlap.
- Vidu voice clone $1.50 and audio $0.05 vs Sume avatar and music
Vidu: voice clone $1.50, text-to-audio under 5 s $0.05. Sume: avatar creation $0.95, Music 1.0 $0.125. Not like-for-like; here is how to budget each.
- Wan 3.0's 20 reference assets vs Omni's 10 images and 3 clips
Alibaba lists up to 20 reference assets for Wan 3.0; Sume's Omni row takes 10 images and 3 clips of 3 s. Which to pick for a branded clip, with costs per 10 s.
- Wan 3.0 takes documents as references. What Sume's route accepts
Alibaba lists documents, webpages and text among Wan 3.0's 20 reference assets. Sume's video route takes image, video and audio references. How to bridge it.
Written by Sume