# AI Video Generation

Source: [AI Video Generation](<https://pr-995-preview.chatcut.dev/docs/video-generation-models>)

Documentation index: [llms.txt](<https://pr-995-preview.chatcut.dev/docs/llms.txt>)

Last updated: 2026-11-19

AI Video Generation creates clips from text, images, first and last frames, references, or a source video. Models support different inputs, editing operations, durations, and resolutions, so define the task before selecting a model.

Video generation uses credits. Seedance, Kling, and Gemini Omni models require a paid plan; depending on your plan, MiniMax H3 and H3 Max may also require an upgrade. If an upgrade is needed, ChatCut says so before generating, and your credits are kept. Use the estimate shown in the editor before sending; see [Credits](https://pr-995-preview.chatcut.dev/docs/credits-policy#how-many-credits-does-video-generation-use) for current rates.

## What do I need before writing the prompt?

-   First decide whether you are generating from scratch, connecting first and last frames, using references, editing a source video, or extending a clip. That choice determines the available models and inputs.
-   Text-only generation needs no upload, but you should know the subject, action, setting, camera, and visual style.
-   To preserve a person, product, or composition, prepare clear images, first and last frames, or supported references. To edit or extend a clip, import the source video into the **Media Pool** first.
-   Decide the aspect ratio, target duration, resolution, and anything that must remain unchanged.
-   You can switch to **Video Gen** for manual setup or stay in Agent mode and describe the task in natural language. When the current Agent supports commands, typing `/` opens its available commands; the list varies by Agent.

## Prompts you can copy

Prompt

```text
Generate a 6-second 16:9 product shot of black wireless headphones floating against a dark background. Slowly push the camera forward, add a soft blue rim light, and use realistic commercial lighting. Do not add text.
```

Prompt

```text
Generate a natural transition from @first-frame to @last-frame. Keep the person's face, clothing, and background consistent. Have the person turn away from the camera and walk out of frame; keep the camera locked.
```

Prompt

```text
Edit @source-video: remove the cup from the desk and replace the background with a clean light-gray studio. Keep the person, action, camera, and original duration unchanged.
```

State what each reference controls and which details may change or must remain fixed.

## Configure Video Gen manually

1.  Switch the AI composer to **Video Gen**.
2.  Select the model name below the composer, then choose **Seedance 2.5**, **Seedance 2.5 Flash**, **Seedance 2.0**, **Seedance 2.0 Fast**, **Seedance 2.0 mini**, **Kling 3.0**, **Kling 4.0**, **Kling 4.0 Flash**, **MiniMax H3**, **MiniMax H3 Max**, or **Gemini Omni 1.1 Flash**.
3.  Select the **Video settings** slider icon. Under **Input**, choose a supported operation, such as **Frames**, **References**, **Edit video**, or **Extend video**. The choices depend on the model.
4.  Set **Aspect ratio**, **Resolution**, and **Duration**. Seedance 2.5 offers 480p, 720p, 1080p, or native 4K, and also MP4 or MOV output.
5.  Upload media through the displayed slots, or drag compatible media from the **Media Pool**.
6.  Enter the prompt, check the credit estimate beside Send, and select **Send message**.
7.  Follow progress in the **Media Pool**. When the video is finished, [download the generated clip](https://pr-995-preview.chatcut.dev/docs/my-assets#how-do-i-download-a-generated-video) or drag it to the timeline to keep editing.

Each successful generation gives you one video. To request another take, submit the same prompt again; each take is a separate job, asset and charge.

## Choose a model

-   **[Seedance 2.5](https://pr-995-preview.chatcut.dev/models/seedance-2.5)**: for mixed references, source-video editing, or clip extension, at 4–30s and up to native [4K](https://pr-995-preview.chatcut.dev/models/seedance-2.5-4k). A 4K clip takes longer and costs about 1.6 times as much as 1080p. A 480p clip is a draft: ask the Agent to turn an approved draft into a matching 1080p or 4K final.
-   **[Seedance 2.5 Flash](https://pr-995-preview.chatcut.dev/models/seedance-2.5-flash)**: for trying a shot faster and for less, at 480p or 720p. It takes the same kinds of input as Seedance 2.5, including editing and extension, with up to 10 images, 5 videos, and 5 audio files as references.
-   **[Seedance 2.0](https://pr-995-preview.chatcut.dev/models/seedance-2.0)**: for standard text, first/last-frame, and mixed-reference generation, up to 4K.
-   **Seedance 2.0 Fast / mini**: for faster or lower-cost iterations, at 480p or 720p.
-   **Gemini Omni 1.1 Flash**: for generation from a first frame or up to 3 reference images, or for editing or extending a source clip of up to 10 seconds; 3–10s clips at 360p to 720p, with 1080p and 4K available as upscaled output.
-   **MiniMax H3**: for multimodal image, video, or audio references, a last frame without a first frame, or optional 2K output.
-   **MiniMax H3 Max**: the faster H3 variant, for text or first-frame generation at 480p or 768p; in **Video Gen** it does not accept reference media, and it has no 2K output.
-   **[Kling 3.0](https://pr-995-preview.chatcut.dev/models/kling-3.0)**: for text, first/last-frame, and reference-image generation in one request (up to 7 images in all), at 3–15s and 720p or 1080p.
-   **[Kling 4.0](https://pr-995-preview.chatcut.dev/models/kling-4.0)**: for the longest single clip (up to 30s) or 21:9 output; combines frames with up to 10 images and 5 reference videos. Name each input in the prompt, such as @image\_1 or @video\_1. Kling 4.0 (not Flash) can also take middle keyframes when you ask the Agent.
-   **Kling 4.0 Flash**: the same inputs as Kling 4.0 for lower-cost 3–20s clips at 720p.

The [Model Capability Matrix](https://pr-995-preview.chatcut.dev/docs/model-capability-matrix) is the source of truth for changing input counts, durations, resolutions, editing, and extension support. Switching models can remove incompatible inputs, so recheck the slots and settings before sending.

## Edit or extend an existing video

With **Seedance 2.5** or **Seedance 2.5 Flash**, choose **Video settings → Input → Edit video** or **Extend video**, add the source clip, and describe the change. With **Gemini Omni 1.1 Flash**, choose **Edit video** or **Extend video** and drag a project video of up to 10 seconds from the Media Pool into the source slot.

When aspect ratio, duration, or resolution must inherit from the source clip, the editor locks or labels that setting. The result is added to the Media Pool as a new asset; it does not overwrite the source file.

See also [Generation References](https://pr-995-preview.chatcut.dev/docs/generation-references) and [Generation Credits](https://pr-995-preview.chatcut.dev/docs/generation-credits).
