Docs

Torch Object Storage

AIGC-Doubao LLMs

Combining TOS and Doubao LLMs for visual understanding, image generation, and video generation

Image processing is in high demand in many fields, including Internet companies, the media, and e-commerce. The Doubao models on Volcano Engine Ark (VeArk) demonstrate strong capabilities in visual understanding, image generation, and video generation. TOS provides robust support for these models with its efficient data storage and flexible processing capabilities. This topic describes how to leverage these capabilities to optimize model pre- and post-processing workflows, thereby improving processing efficiency and quality.

Background information

VeArk supports a variety of models, including:

Visual understanding: Multimodal large language models with powerful image understanding, reasoning, and precise instruction-following capabilities. These models excel at extracting information from images and performing reasoning tasks. They can be applied to a wide range of complex visual question-answering scenarios.

Text-to-image&Image-to-image: These models support various visual styles. With domain-specific services, they can be used in many areas, including education, entertainment, and offices.

Video generation: These models can generate videos from text or images. Their strong semantic understanding capabilities enable them to quickly produce high-quality video clips for film production, advertising, and other applications.

The list of supported models is available in Model list.

Constraints and limitations

Due to model-specific constraints, you may need to do the following when using image-related models:

Operation	Description
Format conversion	Only a limited set of image formats is supported. For example, photos taken on iOS devices are in .heic format and must be converted before they can be used.
Compression	File size limits are enforced. If your image exceeds the limit, you must compress it before using these models.
Resizing	These models convert images into tokens prior to performing inference. The number of tokens used is calculated based on the width and height of the image in pixels. Resizing images allows more images to be processed and reduces costs.
Optimizing data transfer cost	If you need to use a locally stored image, you must convert it to Base64 encoding before submitting it to any of these models. Base64 encoding increases the amount of data transferred over the network.
Video frame capture	Visual understanding models can only interpret image content. When you upload a video, the model must first extract static frames from it and then process them.
Processing of model outputs	Images produced by these models typically require post-processing before they can be delivered to downstream applications. This may include adding annotations or watermarks.

Capabilities

Integrating TOS with VeArk models can significantly improve your experience working with large language models. TOS offers efficient, flexible, and scalable image processing solutions. It supports a variety of image processing capabilities and can be used with visual understanding models, text-to-image models, and video generation models on VeArk. See Model list for a complete list of supported models.

tip

If you would like to use a model that is not on the list, submit a ticket to contact our technical team.

Solutions

We offer the following solutions based on TOS capabilities and business scenarios:

Scenario	Steps
Image preprocessing	As described in Steps 1 and 2, upload the images to be preprocessed to TOS. TOS supports multiple upload methods and storage types to meet your needs in various scenarios. (Optional) Use TOS's image processing features, such as format conversion and compression. Then, use the URL of the processed image as input for VeArk models. tip You can store the preprocessed images in TOS for future reuse. After the inference process is complete, you can retrieve the output directly from VeArk or save it to a TOS bucket under your Volcano Engine account, which is strongly recommended.
Image postprocessing	Image Use a text-to-image model to generate images and save them to a TOS bucket. Use TOS's image processing features, such as adding watermarks and enabling original image protection. After postprocessing, distribute the images to target clients or applications using TOS distribution features, such as sharing, pre-signed URLs, or combining with Content Delivery Network (CDN). Video Use a video generation model to create videos and save them to a TOS bucket. Use TOS's image processing features, such as video frame capture. After postprocessing, distribute the videos to target clients or applications using TOS distribution features, such as sharing, pre-signed URLs, or combining with CDN.

Benefits

This section uses a visual understanding model to demonstrate the benefits of combining TOS and Doubao LLMs.

Category	Combining your own image processing tool and Doubao LLMs	Combining TOS and Doubao LLMs
Size and format restrictions	Maximum image file size: 10 MB. The recommended size is 500 KB or less. Supported image file formats: JPEG, WebP, etc. For a list of supported formats, see Image understanding.	Recommended file size: 20 MB or less. Supported file formats: Image file formats (e.g., JPG, JPEG, PNG, BMP, GIF, WebP, and TIFF) and video file formats. For more information about limitations, see Image processing overview.
Compression method	You need to compress the files yourself using open-source components.	You can use TOS's built-in features to process images directly.
Computing power	This method consumes local computing power.	This method uses cloud-based computing resources with dedicated hardware acceleration.
Performance	-	This method uses TOS for Base64 encoding, which involves less data transfer.
Costs	Data transfer via the public network/Data transfer: CNY 0.5 per GB Requests: CNY 0.01–1 per 10,000 requests Data processing: Local compute costs	Data transfer: Free. Data is transferred via the private network. Image processing: You receive 10 TB of free processing quota. After that, the cost is CNY 0.025 per GB. For more information, see Data processing.
Complexity	You can debug locally, which is more convenient.	You only need to configure it once in the console, and no further action is required. warning Configuration is automated, though initial setup may require attention. Before configuring, confirm which image processing style you need.

Preparation

Preparing TOS

Preparing VeArk

Activate TOS.

Create a bucket and obtain its private network access domain. For steps to create a bucket using the console, see Creating buckets.

tip

Since accessing TOS from VeArk is done over a private network, no fees are incurred for outbound traffic over the public network.

(Optional) Upload the images or videos you want to process to TOS.

TOS supports multiple upload methods. For more information, see Getting Started. If your images or videos are stored elsewhere, you can use Data Migration Service to transfer them to TOS.

The following example uses the Python SDK to demonstrate how to upload local images or videos to TOS. For more information about the Python SDK, see Overview (Python SDK).

import os
import tos

# Get AK and SK from the environment variables.
ak = os.getenv('TOS_ACCESS_KEY')
sk = os.getenv('TOS_SECRET_KEY')
# Fill in the endpoint and region corresponding to the bucket. # For example, fill in "tos-cn-beijing.ivolces.com" (endpoint) and "cn-beijing" (region) for China (Beijing).
endpoint = "your endpoint"
region = "your region"
bucket_name = "bucket-test"
# Object name. For example, fill in "example_dir/example_object.jpg" to upload the "example_object.jpg" object under "example_dir".
object_key = "object-test"
# The path for the local file you want to upload.
file_name = "/usr/local/test.jpg"
try:
# Create a TosClientV2 object. All operations on buckets and objects are performed through this object.
client = tos.TosClientV2(ak, sk, endpoint, region)
# Upload the local file to the bucket.
# file_name is the full path of the local file.
client.put_object_from_file(bucket_name, object_key, file_name)
except tos.exceptions.TosClientError as e:
# The operation failed and a client exception was caught. This is usually due to invalid request parameters or a network exception.
print('fail with client error, message:{}, cause: {}'.format(e.message, e.cause))
except tos.exceptions.TosServerError as e:
# The operation failed and a server exception occurred. See the response for the error details.
print('fail with server error, code: {}'.format(e.code))
# Use the request ID for troubleshooting purposes. We strongly recommend including the request ID in the logs.
print('error with request id: {}'.format(e.request_id))
print('error with message: {}'.format(e.message))
print('error with http code: {}'.format(e.status_code))
print('error with ec: {}'.format(e.ec))
print('error with request url: {}'.format(e.request_url))
except Exception as e:
print('fail with unknown error: {}'.format(e))

Depending on the type of resource you need to process and your output requirements, you must either create an image style or obtain the specific parameters of an existing one. The key features of TOS image processing are explained below. For more features, see Image processing overview.

Format conversion

Compression (resizing)

Compression (quality change)

Watermark (text/image watermark)

Video frame capture

Saving processed images

Original image protection

Description: This feature converts the format of an image.

Documentation: Format

Example:

Parameters: image/format,png

Image processing URL: https://tos-tools.tos-cn-beijing.volces.com/misc/sample.png?x-tos-process=image/format,png

Image style configuration:

Examples

This section provides example codes for different scenarios based on TOS data processing features and VeArk model types.

TOS + Visual understanding model

Online inference

The scenarios are described as follows:

Scenario 1: The processed image or video frame files are not saved to TOS. These files are only used once as model input and cannot be reused. If you need to reuse an image or video, you must reprocess it through TOS. This may result in additional data processing and request fees. For more information, see Billing overview.

Scenario 2: The processed image or video frame files are saved to TOS and can be reused.

Scenario 1

Scenario 2

Batch inference

warning

Before using batch inference, make sure you have created a batch inference client. For more information, see Batch inference.

The scenarios are described as follows:

Scenario 1: The processed image or video frame files are not saved to TOS. These files are only used once as model input and cannot be reused. If you need to reuse an image or video, you must reprocess it through TOS. This may result in additional data processing and request fees. For more information, see Billing overview.

Scenario 2: The processed image or video frame files are saved to TOS and can be reused.

Scenario 1

Scenario 2

TOS + Image generation model

The following example uses the VeArk text-to-image model, Doubao-Seedream 5.0, to show how to generate and save images to a TOS bucket and apply image processing.

tip

TOS does not save the processed images by default. If you need to reuse an image, you must reprocess it through TOS. This may result in additional data processing and request fees. For more information, see Billing overview.

TOS + Video generation model

The following example uses the VeArk video generation model, Doubao-Seedance 2.0, to demonstrate how to save a local image file to a TOS bucket, apply image processing, and use the processed image as model input to create a five-second video.

Last updated: 2026.09.16 10:50:51