tuning-Fine-tune a video generation model

If prompt optimization and official video effects do not meet your needs, fine-tune a Wan image-to-video model to learn specific actions, effects, or styles from your training data. Fine-tuning uses SFT-LoRA (Supervised Fine-Tuning with Low-Rank Adaptation) to train a lightweight adapter on the base model, producing a custom LoRA model that consistently reproduces your target visual effect without detailed prompts.

How it works

The end-to-end workflow has five stages:

HELPCODEESCAPE-json
Upload dataset --> Create training job --> Deploy model --> Generate videos --> (Optional) Evaluate checkpoints
                         |                                                              |
                    Poll until SUCCEEDED                                          Select best and redeploy

1. Upload a ZIP dataset containing training images, videos, and annotations.

2. Train by creating a fine-tuning job. Training takes several hours depending on the model and dataset size.

3. Deploy the fine-tuned model as an online service. This step takes 5--10 minutes.

4. Generate videos by calling the deployed model with an input image.

5. (Optional) Evaluate intermediate checkpoints to find the best-performing model snapshot.

Before you begin

• Region: This feature is available only in the Singapore region under the international deployment mode.

• API key: Get an API key for the Singapore region, then set it as an environment variable:

  HELPCODEESCAPE-bash
    export DASHSCOPE_API_KEY="sk-xxxxxxxxxxxxxxxxxxxxxxxx"

Supported base models

Mode	Base models
Image-to-video (first frame)	wan2.6-i2v, wan2.5-i2v-preview, wan2.2-i2v-flash
Image-to-video (first and last frames)	wan2.2-kf2v-flash

What fine-tuning can and cannot do

Fine-tuning teaches the model new visual content and dynamics but does not change video specifications (resolution, frame rate, duration) -- these are determined by the base model.

Good use cases for fine-tuning:

• Fixed visual effects -- carousels, magic costume changes, money rain

• Fixed character actions -- specific dance moves, martial arts stances

• Fixed camera movements -- push-pull, pan-tilt, orbiting shots

Example: before and after fine-tuning

Image-to-video (first frame): "money rain" effect

Goal: Train a LoRA model that generates a "money rain" video from any input image -- no prompt needed.

Input first frame	Before fine-tuning	After fine-tuning
	Video -- Motion is uncontrollable; the effect varies each time.	Video -- The fine-tuned model consistently reproduces the "money rain" effect from the training set.

Image-to-video (first and last frames): "fashion magazine" effect

Goal: Train a LoRA model that generates a "fashion magazine" transition between a first frame and last frame.

Input first frame	Input last frame	Before fine-tuning	After fine-tuning
		Video -- Motion is uncontrollable.	Video -- Consistently reproduces the "fashion magazine" effect.

Step 1: Upload a dataset

Upload a local .zip dataset to Model Studio to get a file ID.

Sample datasets for the examples above:

• First frame mode: wan-i2v-training-dataset.zip

• First and last frames mode: wan-kf2v-training-dataset.zip

To build your own dataset, see Build a custom dataset.

Request

This example uploads the first-frame training dataset. Only the training set is uploaded -- the system auto-splits a validation set. Upload time varies by file size.

HELPCODEESCAPE-bash
curl --location --request POST 'https://dashscope-intl.aliyuncs.com/compatible-mode/v1/files' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--form 'file=@"./wan-i2v-training-dataset.zip"' \
--form 'purpose="fine-tune"'

Response

Save the id value -- the unique identifier for the uploaded dataset.

HELPCODEESCAPE-json
{
    "id": "file-ft-b2416bacc4d742xxxx",
    "object": "file",
    "bytes": 73310369,
    "filename": "wan-i2v-training-dataset.zip",
    "purpose": "fine-tune",
    "status": "processed",
    "created_at": 1766127125
}

Step 2: Create a fine-tuning job

2.1 Start the job

Use the file ID from Step 1 to start training.

Hyperparameter values vary by model. See Hyperparameters and Request examples for details.

Request

Replace <training_dataset_file_id> with the id from Step 1.

First frame mode

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/fine-tunes' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--header 'Content-Type: application/json' \
--data '{
    "model":"wan2.6-i2v",
    "training_file_ids":[
        "<training_dataset_file_id>"
    ],
    "training_type":"efficient_sft",
    "hyper_parameters":{
        "n_epochs":400,
        "batch_size":2,
        "learning_rate":2e-5,
        "split":0.9,
        "eval_epochs": 50,
        "max_pixels": 36864
    }
}'

First and last frames mode

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/fine-tunes' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--header 'Content-Type: application/json' \
--data '{
    "model":"wan2.2-kf2v-flash",
    "training_file_ids":[
        "<training_dataset_file_id>"
    ],
    "training_type":"efficient_sft",
    "hyper_parameters":{
        "n_epochs":400,
        "batch_size":4,
        "learning_rate":2e-5,
        "split":0.9,
        "eval_epochs": 50,
        "max_pixels": 262144
    }
}'

Response

Key fields in output:

• job_id -- Use this to query training progress

• finetuned_output -- Fine-tuned model name for deployment and invocation

• status -- Initial: PENDING

HELPCODEESCAPE-json
{
    ...
    "output": {
        "job_id": "ft-202511111122-xxxx",
        "status": "PENDING",
        "finetuned_output": "xxxx-ft-202511111122-xxxx",
        ...
    }
}

2.2 Poll the job status

Poll this endpoint with job_id from Step 2.1 until status is SUCCEEDED.

Training takes several hours, varying by model and dataset size.

Request

Replace <job_id> with the value from Step 2.1.

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/fine-tunes/<job_id>' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--header 'Content-Type: application/json'

Response

Key fields in output:

• status -- SUCCEEDED means ready for deployment

• usage -- Total tokens consumed (for billing)

HELPCODEESCAPE-json
{
    ...
    "output": {
        "job_id": "ft-202511111122-xxxx",
        "status": "SUCCEEDED",
        "usage": 432000,
        ...
    }
}

Step 3: Deploy the fine-tuned model

3.1 Create a deployment

Deploy the model as an online service once status is SUCCEEDED.

Request

Replace <model_name> with the finetuned_output value from Step 2.1.

First frame mode

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/deployments' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--header 'Content-Type: application/json' \
--data '{
    "model_name": "<model_name>",
    "aigc_config": {
        "use_input_prompt": false,
        "prompt": "Provide a video description based on the image content. The description must include the phrase \"Then the s86b5p money rain effect begins, countless huge-sized US dollar bills (beige background/dark green patterns) pour down like a torrential rain, densely hitting and surrounding her. The bills continue to fall, she/he stretches her/his arms upward, neck slightly tilted back, expression surprised, completely immersed in this wild money rain.\"\nOutput Template:\nThe video begins with a shot of [subject description]. [Environment description]. Then the s86b5p money rain effect begins, countless huge-sized US dollar bills (beige background/dark green patterns) pour down like a torrential rain, densely hitting and surrounding her. The bills continue to fall, she/he stretches her/his arms upward, neck slightly tilted back, expression surprised, completely immersed in this wild money rain.\nExample:\nThe video begins showing a young woman standing in front of a brick wall covered with ivy. She has long, smooth reddish-brown hair, wearing a white sleeveless dress, a shiny silver necklace, and a smile on her face. The brick wall in the background is covered with green vines, appearing rustic and natural. Then the s86b5p money rain effect begins, countless huge-sized US dollar bills (beige background/dark green patterns) pour down like a torrential rain, densely hitting and surrounding her. The bills continue to fall, she stretches her arms upward, neck slightly tilted back, expression surprised, completely immersed in this wild money rain.",
        "lora_prompt_default": "Then the s86b5p money rain effect begins, countless huge-sized US dollar bills (beige background/dark green patterns) pour down like a torrential rain, densely hitting and surrounding her. The bills continue to fall, she/he stretches her/his arms upward, neck slightly tilted back, expression surprised, completely immersed in this wild money rain."
    },
    "capacity": 1,
    "plan": "lora"
}'

First and last frames mode

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/deployments' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--header 'Content-Type: application/json' \
--data '{
    "model_name": "<model_name>",
    "aigc_config": {
        "use_input_prompt": false,
        "prompt": "Provide a video description based on the image content. The description must include the phrase \"Then she/he begins the s86b5p transformation.\"\nOutput Template:\nThe video begins with a shot of [subject description]. [Environment description]. Then she/he begins the s86b5p transformation.\nExample:\nThe video begins with a young woman in an outdoor setting. She has short, curly dark brown hair and a friendly smile. She is wearing a black Polo shirt with colorful floral embroidery. The background features green vegetation and distant mountains. Then she begins the s86b5p transformation.",
        "lora_prompt_default": "Then she/he begins the s86b5p transformation."
    },
    "capacity": 1,
    "plan": "lora"
}'

Response

Key fields in output:

• deployed_model -- Model name for checking status and invocation

• status -- Initial: PENDING

HELPCODEESCAPE-json
{
    ...
    "output": {
        "deployed_model": "xxxx-ft-202511111122-xxxx",
        "status": "PENDING",
        ...
    }
}

3.2 Poll the deployment status

Poll until status is RUNNING.

Deployment takes 5 to 10 minutes.

Request

Replace <deployed_model> with the value from Step 3.1.

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/deployments/<deployed_model>' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--header 'Content-Type: application/json'

Response

RUNNING status means ready for invocation.

HELPCODEESCAPE-json
{
    ...
    "output": {
        "status": "RUNNING",
        "deployed_model": "xxxx-ft-202511111122-xxxx",
        ...
    }
}

Step 4: Generate videos

Call the model to generate videos once status is RUNNING. 4.1 Submit a video generation taskRequest

Replace <deployed_model> with the deployed_model value from Step 3.

First frame mode

Submits an async task. The model generates "money rain" videos from input images without prompts.

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/services/aigc/video-generation/video-synthesis' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--header 'Content-Type: application/json' \
--header 'X-DashScope-Async: enable' \
--data '{
    "model": "<deployed_model>",
    "input": {
        "img_url": "https://help-static-aliyun-doc.aliyuncs.com/file-manage-files/en-US/20251219/xmvyqn/lora.webp"
    },
    "parameters": {
        "resolution": "720P",
        "prompt_extend": false
    }
}'

First and last frames mode

The model generates a video with the "fashion magazine" effect that transitions between the first and last frames.

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/services/aigc/image2video/video-synthesis' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--header 'Content-Type: application/json' \
--header 'X-DashScope-Async: enable' \
--data '{
    "model": "<deployed_model>",
    "input": {
        "first_frame_url": "https://help-static-aliyun-doc.aliyuncs.com/file-manage-files/zh-CN/20260113/typemn/kf2v-first.webp",
        "last_frame_url": "https://help-static-aliyun-doc.aliyuncs.com/file-manage-files/zh-CN/20260113/ekzmff/kf2v_last.webp"
    },
    "parameters": {
        "resolution": "720P",
        "prompt_extend": false
    }
}'

Response

Save the task_id for querying results.

HELPCODEESCAPE-json
{
    "output": {
        "task_status": "PENDING",
        "task_id": "0385dc79-5ff8-4d82-bcb6-xxxxxx"
    },
    "request_id": "4909100c-7b5a-9f92-bfe5-xxxxxx"
}

LoRA model parameters

Examples use X-DashScope-Async for async tasks. Fine-tuned LoRA models use mostly the same parameters as the standard API. Tables below show LoRA-specific behavior and limitations.

For unlisted parameters (e.g., duration), see the standard API:

• First frame mode: Image-to-video API

• First and last frames mode: Image-to-video (first and last frames) API

First frame mode parameters

Field	Type	Required	Description	Example
model	string	Yes	Name of a fine-tuned model with RUNNING deployment status.	xxxx-ft-202511111122-xxxx
input.prompt	string	No	Text prompt. If aigc_config.use_input_prompt is true, uses this prompt; if false, auto-generates from template.	-
input.img_url	string	Yes	First frame image URL. For supported input methods, see img_url.	https://example.com/image.jpg
parameters.resolution	string	No	Output resolution. wan2.2/2.5: 480P or 720P. wan2.6: 720P or 1080P. Default: 720P.	720P
parameters.prompt_extend	boolean	No	Enable prompt rewriting. Set to false for LoRA models.	false

First and last frames mode parameters

Field	Type	Required	Description	Example
model	string	Yes	Name of a fine-tuned model with RUNNING deployment status.	xxxx-ft-202511111122-xxxx
input.prompt	string	No	Text prompt. Behavior depends on aigc_config.use_input_prompt: if true, the system uses this prompt; if false, the system ignores this field and auto-generates a prompt from the preset template.	-
input.first_frame_url	string	Yes	First frame image URL. For supported input methods, see first_frame_url.	https://example.com/first.jpg
input.last_frame_url	string	No	Last frame image URL. For supported input methods, see last_frame_url.	https://example.com/last.jpg
parameters.resolution	string	No	Output resolution. Fine-tuned models support 480P or 720P. Default: 720P.	720P
parameters.prompt_extend	boolean	No	Enable prompt rewriting. Set to false for LoRA models.	false

4.2 Get the generated video Poll until task_status is SUCCEEDED, then download the video.

Request

Replace <task_id> with the value from Step 4.1.

HELPCODEESCAPE-bash
curl -X GET 'https://dashscope-intl.aliyuncs.com/api/v1/tasks/<task_id>' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY"

ResponseImportant

The video URL expires in 24 hours. Download it before it expires.

HELPCODEESCAPE-json
{
    "request_id": "c87415d2-f436-41c3-9fe8-xxxxxx",
    "output": {
        "task_id": "a017e64c-012b-431a-84fd-xxxxxx",
        "task_status": "SUCCEEDED",
        "submit_time": "2025-11-12 11:03:33.672",
        "scheduled_time": "2025-11-12 11:03:33.699",
        "end_time": "2025-11-12 11:04:07.088",
        "orig_prompt": "",
        "video_url": "https://dashscope-result-sh.oss-cn-shanghai.aliyuncs.com/xxx.mp4?Expires=xxxx"
    },
    "usage": {
        "duration": 5,
        "video_count": 1,
        "SR": 480
    }
}

Build a custom dataset

Build custom datasets for fine-tuning unique effects.

A dataset has a training set (required) and optional validation set . Package as .zip using only English letters, numbers, underscores, or hyphens in filenames.

Training set format

First frame mode

The training set contains a first frame image, training video, and annotation file (data.jsonl).

• Sample: wan-i2v-training-dataset.zip

• Folder structure:

  HELPCODEESCAPE-json
    wan-i2v-training-dataset.zip
    ├── data.jsonl        # Required. Max size: 20 MB.
    ├── image_1.jpeg      # Max resolution: 4096 x 4096. Formats: BMP, JPEG, PNG, WEBP.
    ├── video_1.mp4       # Max resolution: 4096 x 4096. Formats: MP4, MOV.
    ├── image_2.jpeg
    └── video_2.mp4

• Annotation file (data.jsonl): Each line is a JSON object representing one training sample.

  HELPCODEESCAPE-json
    {
        "prompt": "The video begins showing a young woman standing in front of a brick wall covered with ivy. She has long, smooth reddish-brown hair, wearing a white sleeveless dress, a shiny silver necklace, and a smile on her face. The brick wall in the background is covered with green vines, appearing rustic and natural. Then the s86b5p money rain effect begins, countless huge-sized US dollar bills (beige background/dark green patterns) pour down like a torrential rain, densely hitting and surrounding her. The bills continue to fall, she stretches her arms upward, neck slightly tilted back, expression surprised, completely immersed in this wild money rain.",
        "first_frame_path": "image_1.jpg",
        "video_path": "video_1.mp4"
    }

First and last frames mode

The training set contains a first frame image, a last frame image, a training video, and an annotation file (data.jsonl).

• Sample: wan-kf2v-training-dataset.zip

• Folder structure:

  HELPCODEESCAPE-json
    wan-kf2v-training-dataset.zip
    ├── data.jsonl                # Required. Max size: 20 MB.
    ├── image/                    # First and last frame images.
    │   ├── image_1_first.jpg     # Max resolution: 4096 x 4096. Formats: BMP, JPEG, PNG, WEBP.
    │   └── image_1_last.png
    └── video/                    # Training videos.
        ├── video_1.mp4           # Max resolution: 4096 x 4096. Formats: MP4, MOV.
        └── video_2.mov

• Annotation file (data.jsonl):

  HELPCODEESCAPE-json
    {
        "prompt": "The video begins by showing a young woman in an outdoor setting. She has short, curly dark brown hair, a smile on her face, and looks very friendly. She is wearing a black polo shirt with colorful floral embroidery, with a background of green vegetation and distant mountains. Then she begins the s86b5p transformation.",
        "first_frame_path": "image/image_1_first.jpg",
        "last_frame_path": "image/image_1_last.jpg",
        "video_path": "video/video_1.mp4"
    }

Validation set format

Validation set (optional) contains images and data.jsonl -- no videos. The training job generates preview videos from these at each eval step.

First frame mode

• Sample: wan-i2v-valid-dataset.zip

• Folder structure:

  HELPCODEESCAPE-json
    wan-i2v-valid-dataset.zip
    ├── data.jsonl       # Required. Max size: 20 MB.
    ├── image_1.jpeg     # Max resolution: 4096 x 4096. Formats: BMP, JPEG, PNG, WEBP.
    └── image_2.jpeg

• Annotation file (data.jsonl):

  HELPCODEESCAPE-json
    {
        "prompt": "The video begins showing a scene of a young man standing in front of a cityscape. He is wearing a black and white checkered jacket over a black hoodie, with a smile on his face and a confident expression. The background is a city skyline at sunset, with a famous domed building and layered roofs visible in the distance, the sky filled with clouds showing warm orange-yellow hues. Then the s86b5p money rain effect begins, countless huge-sized US dollar bills (beige background/dark green patterns) pour down like a torrential rain, densely hitting and surrounding him. The bills continue to fall while the camera slowly zooms in, he stretches his arms upward, neck slightly tilted back, expression surprised, completely immersed in this wild money rain.",
        "first_frame_path": "image_1.jpg"
    }

First and last frames mode

• Sample: wan-kf2v-valid-dataset.zip

• Folder structure:

  HELPCODEESCAPE-json
    wan-kf2v-valid-dataset.zip
    ├── data.jsonl                 # Required. Max size: 20 MB.
    └── image/                     # First and last frame images.
        ├── image_1_first.jpg      # Max resolution: 4096 x 4096. Formats: BMP, JPEG, PNG, WEBP.
        └── image_1_last.jpg

• Annotation file (data.jsonl):

  HELPCODEESCAPE-json
    {
        "prompt": "The video begins showing a scene of a young man standing in front of a cityscape. He is wearing a black and white checkered jacket over a black hoodie, with a smile on his face and a confident expression. The background is a city skyline at sunset, with a famous domed building and layered roofs visible in the distance, the sky filled with clouds showing warm orange-yellow hues. Then the s86b5p money rain effect begins, countless huge-sized US dollar bills (beige background/dark green patterns) pour down like a torrential rain, densely hitting and surrounding him. The bills continue to fall while the camera slowly zooms in, he stretches his arms upward, neck slightly tilted back, expression surprised, completely immersed in this wild money rain.",
        "first_frame_path": "image/image_1_first.jpg",
        "last_frame_path": "image/image_1_last.jpg"
    }

Data requirements

Requirement	Details
Minimum samples	Minimum: 10. Recommended: 20--100 for stable results.
ZIP package size	≤1 GB (API upload).
Image formats	BMP, JPEG, PNG, WEBP. Max resolution: 4096 x 4096.
Video formats	MP4, MOV. Max resolution: 4096 x 4096.
Video duration	wan2.2: 5 s or less. wan2.5: 10 s or less. wan2.6: 10 s or less.
Individual file size	No limit. System auto-processes files.
Filenames	English letters, numbers, underscores, or hyphens only.

Collect and clean data

1. Acquire raw assets

Choose your method:

Method	Best for	Details
AI generation + curation	Most use cases	Batch-generate videos with Wan base model, then select high-quality samples matching your target.
Real-world footage	Realistic interactions (hugs, handshakes)	Shoot and edit real video clips.
3D rendering	Abstract effects requiring precise control	Use 3D software (Blender, C4D) to create assets.

2. Clean the data

Dimension	Good practice	Common mistake
Consistency	Core features must be consistent. For "360-degree rotation": same direction, same speed across all videos.	Mixed directions -- the model cannot learn which direction is correct.
Diversity	Vary subjects (people, objects), compositions (close-up, long shot, angles), resolution, and aspect ratio.	Single subject/scene -- model may learn irrelevant details (e.g., "red clothes", "white wall") as part of effect.
Balance	If training multiple styles, keep sample counts roughly equal.	90% portrait, 10% landscape -- landscape generation quality suffers.
Purity	Clean visuals without interference.	Watermarks, captions, black bars, or noise -- the model may learn these artifacts as part of the effect.
Duration	Clip assets to ≤ target duration. For 5s videos, use 4--5s clips.	Assets longer than the target duration cause incomplete action learning and truncated results.

Write video annotations (prompts)

Each data.jsonl entry has a prompt describing video content. Prompt quality determines what the model learns.

Prompt formula

Prompt = [Subject description] + [Background description] + [Trigger word] + [Motion description]

Component	Purpose	Required?	Example
Subject description	Describe people or objects in the frame.	Yes	"The video opens with a young woman..."
Background description	Describe the environment.	Yes	"The background is a brick wall covered in green vines..."
Trigger word	A rare, meaningless word that anchors the visual effect.	Recommended	s86b5p or m01aa
Motion description	Describe motion changes in detail.	Recommended	"Enormous US dollar bills pour down like a torrential rain..."

Choose a trigger word

Trigger words act as visual anchors. Many visual effects (unique dance paths, lighting changes, complex transitions) are hard to describe with natural language. Trigger words create direct token-to-effect mappings during training.

How to choose a trigger word:

• Use meaningless letter-number combinations -- s86b5p, sksstyle, a8z2_bbb.

• Avoid common words (beautiful, fire, dance) -- they pollute the model's vocabulary.

Why pair trigger words with motion descriptions? They serve different roles:

• Motion description: basic physical actions (what happens)

• Trigger word: unique visual characteristics words can't capture (how it looks)

Using both together produces the best results.

Keep effect descriptions consistent

Keep motion descriptions identical across all samples with the same effect (training + validation). Only vary subject and background descriptions.

Sample	Prompt
Training sample 1	The video opens with a young woman standing in front of a brick wall... Then the s86b5p money-rain effect begins: enormous US dollar bills (beige background with dark green patterns) pour down like a torrential rain, densely striking and circling her...
Training sample 2	The video opens with a man in a suit inside a high-end restaurant... Then the s86b5p money-rain effect begins: enormous US dollar bills (beige background with dark green patterns) pour down like a torrential rain, densely striking and circling him...
Validation sample 1	The video opens with a young child standing in front of a cityscape... Then the s86b5p money-rain effect begins: enormous US dollar bills (beige background with dark green patterns) pour down like a torrential rain, densely striking and circling him...

Use AI to generate prompts

Use multimodal models like Qwen-VL to draft prompts, then refine manually.

Step 1: Draft descriptions with AI

Send video with a free-form ("Describe the video in detail") or structured template prompt:

HELPCODEESCAPE-python
import os
from openai import OpenAI

client = OpenAI(
    # Get your API key: https://www.alibabacloud.com/help/en/model-studio/get-api-key
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)
completion = client.chat.completions.create(
    model="qwen3-vl-plus",
    messages=[
        {"role": "user", "content": [{
            # When passing a video file directly, set type to video_url
            # OpenAI SDK samples at 1 frame/0.5s (fixed). For custom sampling, use DashScope SDK.
            "type": "video_url",
            "video_url": {"url": "https://cloud.video.taobao.com/vod/Tm1s_RpnvdXfarR12RekQtR66lbYXj1uziPzMmJoPmI.mp4"}​},
            {"type": "text", "text": (
                "Analyze the video carefully and generate a detailed description "
                "using this format:\n"
                "The video opens with [subject description]. "
                "The background is [background description]. "
                "Then the s86b5p melting effect begins: [detailed motion description].\n"
                "Requirements:\n"
                "1. [Subject description]: Describe people or objects in detail -- "
                "appearance, clothing, expressions.\n"
                "2. [Background description]: Describe the environment -- "
                "surroundings, lighting, weather.\n"
                "3. [Motion description]: Describe dynamic changes during the effect -- "
                "object movement, lighting shifts, camera motion.\n"
                "4. Integrate all content naturally. Do not include square brackets."
            )}]
        }]
)
print(completion.choices[0].message.content)

Step 2: Extract and standardize effect templates

Run AI on multiple samples with the same effect. Identify accurate high-frequency phrases, build a template, and apply to all entries.

Keep subject and background unique per sample. Replace only effect descriptions with the template.

Step 3: Review manually

AI may hallucinate. Verify each prompt matches the video (subject, background, motion).

Evaluate models with validation sets

Choose a validation strategy

Training requires a training set. Validation is optional. Two strategies:

Strategy 1: Automatic split (default)

Without validation_file_ids, the system auto-splits training data based on two hyperparameters:

• split -- Training/validation ratio (e.g., 0.9 = 90% training, 10% validation)

• max_split_val_dataset_sample -- Maximum number of samples in the auto-split validation set.

Rule: min(total_samples × (1 - split), max_split_val_dataset_sample)

Example: 100 training samples, split=0.9, max_split_val_dataset_sample=5:

• Theoretical split: 100 x 10% = 10 samples

• Actual validation set: min(10, 5) = 5 samples

Strategy 2: Manual upload Upload custom validation data for checkpoint evaluation.

Providing validation_file_ids disables auto-splitting -- only your uploaded data is used.

Procedure:

1. Package validation data into a .zip file (see Validation set format).

2. Upload the validation set using the upload dataset API to get a file ID.

3. Pass the file ID in validation_file_ids when creating the job:

   HELPCODEESCAPE-json
      {
          "model": "wan2.5-i2v-preview",
          "training_file_ids": ["<training_set_file_id>"],
          "validation_file_ids": ["<validation_set_file_id>"],
          ...
      }

Select the best checkpoint

System saves checkpoints at regular intervals. The last checkpoint becomes the final model by default, but intermediate ones may perform better. Compare to find the best.

Preview videos generate at eval_epochs intervals.

Step 1: View checkpoint preview videos

1.1 List validated checkpoints

Returns only checkpoints with preview videos.

Replace <job_id> with the value from Step 2.1.

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/fine-tunes/<job_id>/validation-results' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY" \
--header 'Content-Type: application/json'

Response:

HELPCODEESCAPE-json
{
    "request_id": "da1310f5-5a21-4e29-99d4-xxxxxx",
    "output": [
        {
            "checkpoint": "checkpoint-160"
        },
        ...
    ]
}

1.2 View a checkpoint's preview videos

Pick a checkpoint from the list (e.g., checkpoint-160) and view its generated videos.

Replace <job_id> and <checkpoint> with the appropriate values.

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/fine-tunes/<job_id>/validation-details/<checkpoint>?page_no=1&page_size=10' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY"

Response:

video_path expires in 24h. Download and review before expiry. Repeat for multiple checkpoints.

HELPCODEESCAPE-json
{
    "request_id": "375b3ad0-d3fa-451f-b629-xxxxxxx",
    "output": {
        "page_no": 1,
        "page_size": 10,
        "total": 1,
        "list": [
            {
                "video_path": "https://finetune-swap-wulanchabu.oss-cn-wulanchabu.aliyuncs.com/xxx.mp4?Expires=xxxx",
                "prompt": "The video begins with a young man sitting in a cafe...",
                "first_frame_path": "https://finetune-swap-wulanchabu.oss-cn-wulanchabu.aliyuncs.com/xxx.jpeg"
            }
        ]
    }
}

Step 2: Export the best checkpoint

2.1 Export the checkpoint

Replace the placeholders:

• <job_id> -- From Step 2.1.

• <checkpoint> -- The checkpoint to export (e.g., checkpoint-160).

• <display_name> -- A custom name for the Model Studio console (e.g., wan2.5-checkpoint-160). Must be globally unique.

For full parameter details, see Export checkpoint.

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/fine-tunes/<job_id>/export/<checkpoint>?model_name=<display_name>' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY"

Response (output=true confirms the export request was created):

HELPCODEESCAPE-json
{
    "request_id": "0817d1ed-b6b6-4383-9650-xxxxx",
    "output": true
}

2.2 Get the exported model name

Query all checkpoint statuses to confirm the export succeeded and get the model_name for deployment.

HELPCODEESCAPE-bash
curl --location 'https://dashscope-intl.aliyuncs.com/api/v1/fine-tunes/<job_id>/checkpoints' \
--header "Authorization: Bearer $DASHSCOPE_API_KEY"

Response:

When the exported checkpoint's status becomes SUCCEEDED , the model_name field contains the name to use for deployment and invocation.

HELPCODEESCAPE-json
{
    "request_id": "b0e33c6e-404b-4524-87ac-xxxxxx",
    "output": [
        ...,
        {
            "create_time": "2025-11-11T13:27:29",
            "full_name": "ft-202511111122-496e-checkpoint-160",
            "job_id": "ft-202511111122-496e",
            "checkpoint": "checkpoint-160",
            "model_name": "xxxx-ft-202511111122-xxxx-c160",
            "model_display_name": "xxxx-ft-202511111122-xxxx",
            "status": "SUCCEEDED"
        },
        ...
    ]
}

Step 3: Deploy and call the exported model

After successfully exporting the checkpoint:

• Deploy: Use the model_name from Step 2.2 as the model_name input parameter (see Step 3: Deploy the fine-tuned model).

• Generate videos: Call the deployed model following Step 4: Generate videos.

Optimize for production

For distorted output, weak effects, or inaccurate motion, try these optimizations.

Common mistakes to avoid

Mistake	Impact	Fix
Inconsistent training data	Model cannot learn the target effect	Make sure all samples show the same effect direction, speed, and style
Too few samples	Weak or unstable effect reproduction	Add at least 20 high-quality samples
Common words as trigger words	Pollutes the model's existing vocabulary	Use meaningless combinations like s86b5p, not real words like running or dance
Assets longer than target duration	Incomplete action learning, truncated output	Clip assets to match the target output duration
Ignoring validation output	Missing the best checkpoint	Monitor preview videos at each checkpoint

Tune hyperparameters

For full parameter descriptions, see Hyperparameters.

• n_epochs: Default 400 . Only change if needed. If adjusted, ensure ≥800 total training steps . Total steps: n_epochs × ceil(training_size / batch_size). Minimum: n_epochs = 800 / ceil(dataset_size / batch_size). Example: 5 training samples, wan2.5 model (batch_size=2):

  • Steps per epoch: ceil(5 / 2) = 3

  • Minimum n_epochs: 800 / 3 = 267 (round up to 300 as a practical minimum)

• learning_rate and batch_size: Use defaults (rarely need changes).

Billing

Item	Billed?	Details
Model training	Yes	Cost = tokens × unit price. See model training billing. Check usage field in query job status for consumption.
Model deployment	No	Free.
Model invocation	Yes	Billed at base model's standard invocation price. See model pricing.

API reference

Video generation model fine-tuning API reference

FAQ

How do I calculate the training and validation set sizes?

Training: required. Validation: optional.

• No validation uploaded: System auto-splits training data.

  • Validation set size = min(total_samples x (1 - split), max_split_val_dataset_sample). See Choose a validation strategy for an example.

  • Training set size = total_samples - validation_set_size.

• Validation uploaded: The system does not split. Training and validation sizes equal the uploaded counts.

How do I design a good trigger word?

• Use meaningless combinations (e.g., s86b5p).

• Avoid common words (e.g., fire).

Can fine-tuning change video resolution or duration?

No. Fine-tuning teaches content and dynamics only. Output specs (resolution, frame rate, duration) are base-model-determined.