Video generation references: roles, counts and input modes

As reviewed on 28 October 2026, reference media guide a generation; they do not automatically enter your edit or guarantee identity. Choose the exact model and one compatible input mode, assign each reference a job, and check both count and aggregate duration. Seedance 2.0 audio needs an image or video; 2.5 has a separate audio-only mode. Max reference input remains blocked in the inspected ChatCut build.

What has actually been checked?

Official specifications were accessed on 27 October, with pinned integration and saved route evidence reviewed 28 October 2026. Local controls do not prove a provider accepted every boundary. The separate local previs example below used one guide and seven images; it is not proof of maximum-count entitlement or current production availability.

Choose the reference path

• Agent request: select + for files, or use Selection Mode to add @ references to project assets, clips, time, Viewer regions or Transcript text. Explain what each selected item means.
• Video Gen: open Video settings → Input and fill the visible first-frame, last-frame, reference or source-video slots. Prompt wording alone does not place a file in a frame slot.
• Existing project media: use compatible assets from My Assets rather than uploading duplicates. An eligible project video or public resource can be a reference; it does not have to be a completed generation.

A prompt attachment and persistent project media have different lifetimes and roles. See My Assets. A Viewer region selected for the Agent does not by itself establish a provider brush/mask input.

Seedance: counts, durations and modes

For full 2.0, Fast and Mini, choose text, first frame, first+last frames, or reference mode. A last frame requires a first; frames cannot mix with reference arrays. The reference budget is 9 images, 3 videos, 3 audios. Each video/audio is 2–15 seconds and the aggregate is at most 15 seconds per type. Audio references require an image or video reference. Fast/Mini have the same input rules but different output tiers; see their family reference.

For 2.5, the documented/local caps are 30 images, 10 videos, 10 audios, 50 total. Each video/audio is 2–30 seconds, with 30 seconds total per type. Audio-only references are documented separately from 2.0. Frames and reference mode still cannot mix. Edit sources have a separate 4–30-second requirement; do not apply that minimum to every reference clip. Maximum-boundary supplier acceptance remains unverified after Ark timeouts.

Both native specifications list images below 30 MB, videos up to 200 MB, audio up to 15 MB, plus a 64 MB aggregate request budget. ChatCut’s inspected public-resource path separately caps each image at 5 MiB and video/audio at 50 MiB. Satisfy every applicable limit; ten valid-length videos can still exceed an aggregate duration or request-size budget. Official task API.

MiniMax and Omni use different reference contracts

H3 exposes up to 9 images, 3 videos and 3 audio files, 12 total. Native video/audio are each 2–15 seconds, at most 15 seconds total per type; native images are up to 30 MB, videos 50 MB and audio 15 MB, with a 64 MB request budget. Frame and reference modes are exclusive. H3’s documented last-only/audio-only modes have no completed endpoint generation in this review.

H3 Max blocks all references in the inspected ChatCut integration. Its configured endpoint explicitly requires first with last, and rejects audio-only references despite broader native documentation. A rejection for ten images or four videos establishes a validator, not successful in-limit generation. See H3 and Max.

Gemini Omni is gemini-omni-1.1-flash-preview, not Veo or Kling. Its pinned path exposes up to three image references or frames and a separate short source-video path; do not import Seedance counts or audio arrays. See the Omni reference for its 10-second source guide and measured encoding tolerance.

Assign identity, camera and sound separately

Use an explicit role statement: “Image 1 controls the product silhouette and color; video 1 controls camera path only; image 2 controls the set and light. Do not copy the guide’s proxy bodies or text.” State what may change. A camera guide, subject sheet and mood board can conflict; adding more files does not resolve that conflict.

In a 27 October local Seedance 2.5 previs run, the submitted request used one 30-second guide and seven images, requested 30 seconds, 21:9 and 1080p. The prompt assigned camera/cuts/blocking to the guide, six identity sheets to appearance and the seventh image to the diner. A 30-second export completed with 15/17 checks passing. Side-view proof failed; speaker assignment and no-music audio remained unverified. This supports one smaller-set workflow, not the maximum reference count or guaranteed identity preservation.

Why did an input disappear or get cropped?

Switching models or Generate/Edit/Extend can remove incompatible slots. Recheck every input, resolution, duration and estimate. A fit/crop or output-ratio choice can change composition; inspect the preview and returned frames before assuming a reference was lost. Missing subject detail may also be a generation miss: a completed job does not prove identity or exact camera reproduction. Use workflow failures to separate validation, unavailable routes and output-quality problems.

Last verified: 28 October 2026. Public sources: Seedance task API, MiniMax V2 task API, accessed 27 October. Pinned tool handling, saved route checks and the separately scoped 27 October local previs report were reviewed 28 October. See model matrix, video generation and Credits Policy before a new request.