AI video generation: model selection, prompts, edit and extend

As reviewed on 28 October 2026, ChatCut’s inspected Video Gen controls offer Seedance, Gemini Omni and MiniMax H3 variants for different generation tasks. Choose the exact model, then a compatible input mode and output settings. Native documentation, visible controls and successful endpoint generation are different evidence. Seedance Draft-to-final is not exposed by the inspected tool; H3 Max references are blocked, and configured Ark boundary checks remain unverified after timeouts.

Models and controls: what is established?

The table summarizes official specifications accessed 27 October and pinned integration/route checks reviewed 28 October 2026. It does not certify the deployed production picker or your account entitlement. Use the options and estimate actually shown in your app; current rates belong to Credits Policy.

Configure Video Gen manually

Kling 3.0 is absent from the inspected manual picker and its new-generation route is suspended. Reinstalling or buying another plan does not create that selector. Choose among the options actually shown; a model’s presence in documentation is not production access proof.

1. Switch the AI composer to Video Gen and choose the exact model.
2. Open the Video settings slider. Under Input, select text generation, frames, references, edit or extend only when exposed for that model.
3. Set Aspect ratio, Resolution and Duration. Seedance 2.5 exposes MP4/MOV output; source-dependent settings may be adaptive or locked.
4. Fill the displayed slots with compatible media, preferably from My Assets. Recheck them after switching models or modes.
5. Enter the prompt, review the model/settings and current credit estimate, then select Send message.
6. Follow the task in My Assets. After completion, inspect the result, download the clip or drag it onto the timeline.

Each new take is a separate job and charge. A completed video is editable as timeline media for trim, crop, captions and assembly; its generated objects are not separate layers. The inspected access rule requires paid video-generation entitlement; check your current account rather than treating a different model as a free workaround.

Write prompts for the selected version

Describe subject → action → setting → camera → sound → constraints. If you add media, assign each a role: identity/wardrobe, set/light, camera path, motion or sound. A first-frame slot anchors an input frame; saying “reference” asks for guidance and does not guarantee an exact composition. See generation references.

For Seedance 2.0/Fast/Mini, write ordered Shot 1, Shot 2 instructions. Timestamp syntax is not guaranteed timing control. 2.5 documents timestamp following, but exact timing on your route still needs checking. The native recommendation is at most 1,000 English words or 500 Chinese characters, not a verified hard rejection limit. “8K cinematic” describes a requested look; the resolution control determines the requested output tier.

Unexecuted single-shot example:

Use image 1 only for the red backpack's shape, fabric and zipper placement.
Shot 1: A hand places the backpack on a plain gray table.
The camera stays locked in a medium close-up. Soft light from the left.
Keep the backpack red and the zipper visible. No text. One continuous take.

A 27 October local Seedance 2.5 previs run submitted this role instruction: “Each of @Image1 to @Image6 is an identity sheet: the face close-up gives the face, the body views give build and clothing.” It used seven images and one 30-second guide, requested 30 seconds, 21:9, 1080p. The raw take was 2206×946 and the final timeline export 2520×1080, 30 seconds. It passed 15/17 checks; useful side-view proof failed and speaker/no-music sound remained unverified. This is one local result, not general timing, identity or resolution proof.

Review a failed shot before paying for another

Separate request errors from creative misses. A local mode/count error needs corrected inputs; a timeout leaves endpoint capability unverified. A completed job with the wrong angle or extra subject is an output miss. Inspect start/middle/end and every cut; audition sound rather than inferring “no music” from a transcript.

A separate 27 October Seedance 2.0 scene run used three image references, 15-second 1080p takes, then two five-second framing repairs. Its final 14.625-second export passed 6/7 MUST checks but retained the wrong opening viewpoint/cant. Stronger angle wording did not fix it. Keep accepted shots, simplify conflicting instructions, revise one failed constraint, and compare the actual next take. Repeating a prompt does not guarantee identical frames.

Is Seedance 2.5 Draft a separate model?

No. Native 2.5 documentation describes a 480p preview → 1080p final workflow using the same model and draft task ID. The draft is retained for 7 days. Finalization cannot re-specify the prompt, references, ratio, duration, seed or audio; changed creative requirements need another draft. The final is similar, not frame-identical, and remains subject to moderation. Seedance 2.5 guide.

The inspected ChatCut tool has no draft or draft-task input and does not forward those fields. This page therefore gives no ChatCut Draft button, expiry guarantee or completed draft-to-final workflow. Mini is an ordinary separate model; Omni 360p is an ordinary low-resolution generation. Neither is the native Seedance draft task.

To compare costs if a route later exposes Draft, obtain matched quotes for ordinary 720p versus preview plus final at the same model, cohort, ratio, duration and references. Add both charges and allow for a discarded preview; no current saving or numerical tariff is established here.

Edit, extend or regenerate?

Edit requests a change to an existing source; extend requests new continuation; generate creates another take. Native Seedance 2.5 edit sources are 4–30 seconds. Edit requires adaptive ratio and automatic duration (-1); extend requires adaptive ratio and describes the added duration. MP4/MOV are the exposed output formats; do not assume they establish every possible source codec. Task API.

The inspected ChatCut tool exposes taskMode=edit/extend. Edit requires exactly one source and prefixes the request with “Edit the video:”; extension prefixes “Extend the video:”. The saved adapter does not forward the native explicit task selector, so those local modes do not prove a native edit/extend endpoint was invoked successfully. Ark checks timed out; a complete region-edit or extension result was not established. TaskTypeMismatch or TaskTypeConstraint can appear after task creation when prompt intent and task constraints disagree.

For an available edit path, use a compatible source in My Assets, choose Edit video, and ask for one change while naming the subject, motion and framing to preserve. Choose Extend video only when that mode is shown and request the new continuation. Review all inherited settings. An Agent Viewer-region reference does not prove a provider region-mask/brush control; no completed region repair is established by this evidence.

What changes between H3 and Max?

Use H3 for the exposed reference mode or 2K tier. Max offers 480p/768p text/frame controls; it does not establish superior fidelity. Native H3 last-only/audio-only modes remain unverified by a completed endpoint job. The Max endpoint explicitly rejected them, and ChatCut blocks all Max references. Give each H3 reference a role and keep frames separate from reference mode.

Gemini Omni is the distinct gemini-omni-1.1-flash-preview: 3–10-second controls, short-source editing/extension and upscaled 1080p/4K labels. No new successful Omni generation was established. Full Seedance 2.0’s native 4K is different: ChatCut blocks it pending pricing, while the configured Ark probe timed out. Do not confuse generation, upscale and export.

Related docs and sources

Use the model matrix, generation references, generation credits and workflow failures. The Seedance prompt guide adds worked formulas and bounded submitted-prompt examples.

Last verified: 28 October 2026. Native sources: Seedance API, 2.5 guide, MiniMax V2, accessed 27 October; pinned controls, endpoint evidence and 27 October local previs/scene reports reviewed 28 October. No fresh generation or production certification was performed for this tutorial.