> For the complete documentation index, see [llms.txt](https://docs.seaart.ai/guide-1/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.seaart.ai/guide-1/advanced-guide/lora-training/video-training.md).

# Video Training

Train a video LoRA to learn a motion or visual effect from a small collection of clips.

This guide covers preparing the videos, describing the action, configuring training, and testing Wan2.1, Hunyuan, and Wan2.2 models.

### Video Training Type Recommendation <a href="#video-training--video-training-type-recommendation" id="video-training--video-training-type-recommendation"></a>

<table><thead><tr><th width="240">Name</th><th>Model</th></tr></thead><tbody><tr><td>t2v (text to video)</td><td>wan2.2 t2v 720p</td></tr><tr><td>i2v (image to video)</td><td>wan2.2 i2v 480p</td></tr><tr><td>r2v (reference to video)</td><td>minimax h3</td></tr></tbody></table>

***

### Wan2.1, Hunyuan Video LoRA Training <a href="#video-training--wan2-1-hunyuan-video-lora-training" id="video-training--wan2-1-hunyuan-video-lora-training"></a>

Start by preparing and tagging the clips. The following sections explain the shared preparation process, the Wan2.1 training settings, and the testing routes for Wan2.1 and Hunyuan.

#### Preprocessing Videos <a href="#video-training--preprocessing-videos" id="video-training--preprocessing-videos"></a>

**Selection of training videos**

* Video content is unified, action or effects are unified, subjects differ. Prefer videos; images can be used as a supplement.
* High-definition quality and no watermark.

**Number of videos**

* 4-10 videos is enough. (Directly using images for training is not recommended.)

**Frame rate**

* Process videos to 16fps, with a total length of 81 frames, 5s in total (you can use editing tools to cut the video to 5s, then extract frames to 16 fps).
* If the video is shorter than 5s, use its actual length, for example 2s or 3s are all fine, but it must be processed to 16fps.

**Resolution**

* Around 480p is enough. You can also process it to 320p to speed up training. (Too high a resolution will make training fail directly.)

**Video tagging**

* Auto tagging (auto annotation is not recommended).

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FA40NQbA910Tz9x1fKzSJ%2FAuto%20tagging.png?alt=media&amp;token=a0f55cca-7c9a-447b-bb88-4941a90d0d88" alt=""><figcaption></figcaption></figure>

* Manual tagging.
* Key points: secondary features + primary features.
* Primary features: the action or effect that needs to be learned. Secondary features: the video's character, where they are, and what they do.

> **Example:**&#x20;
>
> In the video, a <mark style="background-color:orange;">woman wearing a black formal suit</mark> is presented. The person raises <mark style="background-color:orange;">h</mark><mark style="background-color:pink;">er</mark> hand and showers colorful confetti in celebration with smile. The person then reveals a bikini, causing a b1k1n1 bikini up effect. <mark style="background-color:orange;">The person continues celebrating, further showing the b1k1n1 bikini up effect.</mark>

The text before the highlighted part describes the video content, while the highlighted part summarizes the action or effect to be learned. In other words, the highlighted part is the primary feature and the rest are secondary features.

#### Online Parameter Settings <a href="#video-training--online-parameter-settings" id="video-training--online-parameter-settings"></a>

**Image to video**

* Image to video: Wan2.1-14B-480P or Wan2.1-14B-720P, chosen mainly by the resolution of your training video.
* If your material is 216\*320, below 480p, choose the 480p model. The 720p and 480p models train to nearly the same result, so 480p is the recommended starting point.

| Resolution | Specific size | Total pixels |
| ---------- | ------------- | ------------ |
| 480p       | 854\*480      | \~410k       |
| 720p       | 1280\*720     | \~920k       |

Finish uploading the dataset.

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FEdoMYQeDrctZ249dYRwl%2FFinish%20uploading%20the%20dataset..png?alt=media&amp;token=a2ed99ec-39be-4361-88c4-ec45e240db56" alt=""><figcaption></figcaption></figure>

**Parameter Settings**

* Frame extraction rate: how many frames are extracted within a single video segment. Recommended value: 17.
* Example: if a segment is 16 frames, a frame extraction rate of 9 will not capture every frame.
* Slice count: how many segments a single video is split into. Recommended value: 5.
* Example: for a 5-second video at 16 frames per second, a slice count of 5 gives 16 frames per segment, while a slice count of 4 gives 20 frames per segment.
* Repeats per image: how many times each video is learned.
* Cycles: how many times to repeat the process on top of repeats per image.
* Model effect preview prompt: the prompt used to generate the sample video, adjusted to combine the dataset tags with the content of the initial frame.
* Initial frame: for image to video, the image the model needs, and the image used to generate the sample video.

<div align="left"><figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FRMt8hlkLROZimjwjFJP9%2F469c2b08-9bf8-4c38-b051-19a6a9488fd2.png?alt=media&amp;token=812b9e3a-70e8-4803-8a2b-26e0b4c7120d" alt=""><figcaption></figcaption></figure></div>

<div align="left"><figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FlpMMsSI3hcpR60a9W4Wa%2F6d39dcfb-f072-46c6-8cc8-976784fd6e59.png?alt=media&amp;token=f7dc4b95-df78-4d40-9508-f6d13bd9f3b6" alt="" width="360"><figcaption></figcaption></figure></div>

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FHlg3mFJncHLiFZPMToeB%2F42035a33-bf25-4274-96f7-6bf89303b548.png?alt=media&amp;token=8eaa2d98-0159-415b-bd44-cbbb3129a20a" alt=""><figcaption><p>Parameter settings</p></figcaption></figure>

#### Advanced Parameter Settings <a href="#video-training--advanced-parameter-settings" id="video-training--advanced-parameter-settings"></a>

The only value you need to change is **Flow Shift**.

<mark style="color:$danger;">720p takes 5, and 480p takes 3 (the material must also be 480p).</mark>

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2F1oS0U5Scp3wUhZSOvuZ3%2F9531a336-dcfb-488f-9e7c-0256c350a0f2.png?alt=media&amp;token=0011d320-8900-4cae-a77d-3b5a82492058" alt=""><figcaption></figcaption></figure>

### Wan2.1 Testing <a href="#video-training--wan2-1-testing" id="video-training--wan2-1-testing"></a>

* **Image to video**

Workflow:

{% file src="/files/TRfjqzxtTgMq0VfAcxGd" %}

App testing: [SeaArt AI | Wanxiang Guanliu](https://www.seaart.ai/workFlowAppDetail/d1e017de878c73eofvrg)

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FftxtiEEz5rPcSdkPYlDV%2F5b53299b-31f5-47ea-b017-fe0fa1937475.png?alt=media&amp;token=a7cab98c-96db-4564-a833-5e04516233ec" alt=""><figcaption></figcaption></figure>

**Parameter settings:**

* Model selection: use the same model for testing that you used for training.
* Select Lora: choose the Lora saved in your own models.
* Weight: the weight of the Lora.
* Width: the width of the input image after compression and cropping.
* Height: the height of the input image after compression and cropping.
* Frame count: the total frame rate across the output duration, given by 4\*n+1, where n is the number of seconds. Five seconds is 81 frames.
* Shift: 720p is 5, 480p is 3.
* cfg: defaults to 6, and can be adjusted to 5.

**Text to Video Testing**

Wan workflow testing

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2F1ZYNWiNufeS95mfa2hEz%2F5459ab85-5918-4d66-859e-36f26b485dd2.png?alt=media&amp;token=ac2a20be-487b-493b-974d-ff27941bba0c" alt=""><figcaption></figcaption></figure>

[RedCraft Models as a Service H3 GS - SeaArt AI Model](https://www.seaart.ai/models/detail/e68fd840d50f39999fb5aa4d93cf5bf0)

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FGT4IgdJMhHQjsTPO8NIR%2F9d15b098-9af9-4ba9-91e7-726d2afe723d.png?alt=media&amp;token=0094ec0f-08db-45ce-8bf9-5229c9a81ae7" alt=""><figcaption></figcaption></figure>

1. In **Model**, select **wan2.1**.
2. In **Additional**, select the model saved after training.
3. Select **text to video**.

### **Test the Hunyuan LoRA**

Use the Hunyuan workflow testing route and the model page named&#x20;

[Hunyuan Video (Safetensors) - New Decensored Llama - SeaArt AI Model.](https://www.seaart.ai/models/detail/1dcfcdecf715bfe5146cb50a857e16aa)

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2F3Ut6crZT2BRRemmEmzvh%2Fed8af635-9e5b-45e7-938f-73c7f1c4f8ab.png?alt=media&amp;token=51515e18-162a-4b76-993e-307aacf1d0d5" alt=""><figcaption></figcaption></figure>

1. Set **Model** to **Hunyuan Video**.
2. In **Additional**, select the model saved after training.
3. Select **text to video**.

### Wan2.2 Video LoRA Training <a href="#video-training--wan2-2-video-lora-training" id="video-training--wan2-2-video-lora-training"></a>

**Prepare the Videos**

<mark style="background-color:red;">Use the same preparation process as for Wan2.1:</mark>&#x20;

Prepare 4 to 10 clips, process them to 16fps, use the duration and frame guidance above, and keep the resolution around 480p or reduce it to 320p. The video count, duration, frame rate, and resolution requirements are the same.

**Choose the High-Noise and Low-Noise Models**

Each video type has two model choices:

* **Text-to-video:** `wan2.2 t2v-low` and `wan2.2 t2v-high`.
* **Image-to-video:** `wan2.2 i2v-low` and `wan2.2 i2v-high`.

Unlike Wan2.1, Wan2.2 provides a high-noise and a low-noise model for each video type. The high-noise model mainly controls motion, while the low-noise model mainly controls visual detail.

If you want to minimize training time, train only the low-noise model to reach a basic result more quickly. For the best results, train the high-noise and low-noise models separately using the same dataset, then load both LoRAs together when generating a video.

Wan2.2 understands language more effectively, so its video annotations can use simpler, more consistent wording. The appropriate level of detail differs between text-to-video and image-to-video, as explained next.

#### **Tag the Videos for Your Training Type**

**Text-to-video: describe the content clearly**

You can use automatic tagging for text-to-video training.

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FcHjUByEewT148d9ebCR1%2Fautomatic%20tagging%20for%20text-to-video%20training.png?alt=media&amp;token=72626a8c-bcde-46bc-b64b-8aea09764846" alt=""><figcaption></figcaption></figure>

If you write the tags manually, describe the video content clearly. Avoid reducing the whole description to a vague label such as “a person” or “an animal.”

**Image-to-video: keep the effect description simple and consistent**

Automatic tagging is not recommended for this setup. The dataset contains relatively few videos, and automatic tags tend to describe too much detail. With Wan2.2, those detailed tags can make the prompts cumbersome when you later use the LoRA.

For manual tagging, a simple description is enough if all the clips show people.

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FwYt8pjNQUNbonzxEpU1Q%2F886fccca-57a6-45bf-bf88-630ae9b08655.gif?alt=media&amp;token=15a29603-784c-40cb-bc18-0bdc7dec3a07" alt=""><figcaption></figcaption></figure>

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FLZahCkXbw35frkipUCip%2F9172477e-5d29-494b-8e42-f33416e364e6.png?alt=media&amp;token=0b93df1e-5e2b-4fbf-8e16-10a5d0b4c2a4" alt=""><figcaption></figcaption></figure>

For example, the transformation in the guide can be described as:

> <mark style="background-color:yellow;">A person's head turns into a pumpkin, and the person then wears a robe. The Halloween background includes a pumpkin lantern, bats, and the moon.</mark>

You do not need to specify the person's age or gender. Use “a person” to keep the subject description general, and describe the transformation accurately and consistently. You can then copy the sentence to the other video tags.

#### **Configure the Wan2.2 Training Parameters**

Use the following settings for extracting video segments, repeating the training material, and generating previews:

**Frame extraction rate**

This applies within each segment. The recommended value is **17**; a value of 9 for a 16-frame segment would not allow every frame to be learned.

**Slice count**

Set the number of segments created from each video. The recommended value is **5**. For a 5-second video at 16fps, the example gives 16 frames per segment with 5 slices, or 20 frames per segment with 4 slices.

**Repeats per image**

This determines how many times each video is used for training.

**Cycles**

This repeats the process for the specified number of rounds.

**Model effect preview prompt**

Use a prompt based on the dataset tags, adjusted to match the initial-frame image when generating an image-to-video preview.

**Initial frame**

For image-to-video, this is the input image used to generate the sample video.

#### **Set Up Image-to-Video Training**

For a typical image-to-video run, select **wan2.2 i2v-low**. Leave the advanced settings at their default values.

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2F7W6Tv9H6n2nqkGxj6psC%2FSet%20Up%20Image-to-Video%20Training.png?alt=media&amp;token=1dd50494-b686-4809-a799-0134160dda3e" alt=""><figcaption></figcaption></figure>

For **Model Effect Preview Prompts**, you can select the video tags to fill the field directly.

For **Initial Frame**, upload an image that matches the type of your training clips. (Use a realistic image for realistic footage. If the footage shows a half-body view, choose a half-body image for the preview as well.)

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FSYyv6a0AxG5cdb18oCTd%2F333315fe-8112-4711-a993-c73d7ddc3742.png?alt=media&amp;token=c4b965d3-7713-49ba-9954-d230925baed8" alt=""><figcaption></figcaption></figure>

#### **Set Up Text-to-Video Training**

Select **wan2.2 t2v-low**. You can fill **Model Effect Preview Prompts** directly by selecting the video tags.

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FTLR2iuWZZ8CbVGWWfdmY%2FSet%20Up%20Text-to-Video%20Training.png?alt=media&amp;token=f04b5277-4be0-43d8-aa01-14dd59b73bfe" alt=""><figcaption></figcaption></figure>

#### **Follow the Training Demonstrations**

The demonstrations below show the training process for each video type. Follow the one that matches the model you selected.

**Text-to-video (t2v)**

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2F4W0shNfiizEXHlNVf8It%2FText-to-video%20(t2v).gif?alt=media&amp;token=45ecd358-7c81-4105-ba2f-4388cfb17d2f" alt=""><figcaption></figcaption></figure>

**Image-to-video (i2v)**

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2Fcl1RWTxf08dnGCmv6jSr%2FImage-to-video%20(i2v).gif?alt=media&amp;token=d4dba94d-bc49-4a0b-9dcf-3b5b74e58d24" alt=""><figcaption></figcaption></figure>

#### **Test the Wan2.2 LoRA**

**Image-to-video testing**

Use the workflow:

{% file src="/files/ao2wo2uUdncCCTdSHJ5P" %}

or the app [Wan2.2 Test](https://www.seaart.ai/app/d4ihmble878c738mrk2g)

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FENTiGwA5i2OKuDWoeKne%2Fwan%202.2testing.png?alt=media&amp;token=329cb81d-3667-4a3a-8fe9-fb54f91b7d22" alt=""><figcaption></figcaption></figure>

1. Select the **high-noise LoRA** in the first loader.
2. Select the **low-noise LoRA** in the second loader.
3. If you trained only one of them, select that same LoRA in both loaders.
4. **Enter the prompt.** In **Please enter text**, enter the tagging sentence.
5. **Provide an input image.** In **Please select an image**, choose an image, preferably one similar to the first frame of the training video.

**Text-to-video testing**

Use the workflow:

{% file src="/files/DPXzQ8afvgfO97gk69ih" %}

or the app [Wan2.2 Text-to-Video Test](https://www.seaart.ai/app/d4ilp4le878c738nnlg0)

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FwqUCChOvlIPhqSRiSk1Y%2Fwan%202.2textto%20video%20testing.png?alt=media&amp;token=06ea8c1d-a80c-4df9-9297-e323d7811d5a" alt=""><figcaption></figcaption></figure>

1. Select the **high-noise LoRA** in the first loader.
2. Select the **low-noise LoRA** in the second loader.
3. If you have only one LoRA, select it in both loaders.
4. Enter the output video's width in **Please enter the width**.
5. Enter the output video's height in **Please enter the height**.
6. **Enter the prompt.** In **Please enter text**, enter the tagging sentence.
