Building E-commerce Videos with Coze Workflows: A Hands-On Tutorial with Seedance and Happy Horse

Overview

Using the Coze workflow platform combined with Seedance 2.0 or Happy Horse video generation plugins, you can quickly build an automated pipeline for creating e-commerce product showcase videos. This article uses a cat food advertisement as an example to walk through every step from workflow creation to video generation.

Coze Workflow Architecture Overview

Coze is an AI application development platform by ByteDance that allows users to chain together LLMs, plugins, code, and other components through a visual workflow editor to build automated AI applications. Its core advantages are a low-code barrier to entry and a rich plugin ecosystem, enabling non-developers to quickly build complex AI processing pipelines.

The overall workflow logic is straightforward: Start Node (input parameters) → LLM generates video prompt → Video generation plugin → Loop to query results → Output video link.

There are only two core stages: prompt generation and video generation. The subsequent task querying is optional, since you can also look up results directly on the web interface using the task ID.

Start Node Configuration

The start node requires four input variables:

• Video duration: Integer type; Seedance supports 4–15 seconds
• Brand name: String type, e.g., the cat food brand name
• Aspect ratio: String type, e.g., 16:9 or 9:16
• API key: String type; required for calling third-party plugins

Pay attention to variable type selection — duration should be set as an integer to restrict input to numbers only. If you have a real product photo, you can also add an image-type reference image variable.

LLM Prompt Generation Configuration

Model Selection and Parameter Settings

Doubao 2.0 mini is recommended. Key configurations include:

• Disable deep thinking: The task is simple, and disabling this significantly improves execution speed
• Temperature setting: Between 0.3–0.5, balancing precision with moderate creativity
• Max reply length: Set it generously to prevent output truncation errors

Regarding the temperature parameter — it's a core sampling parameter during LLM inference that controls output randomness. Lower values (e.g., 0.1–0.3) produce more deterministic and conservative outputs; higher values (e.g., 0.7–1.0) yield more diverse and creative outputs but may also introduce incoherent content. For structured prompt generation tasks, 0.3–0.5 is a good sweet spot that ensures stable output formatting while allowing some variation in ad creativity.

The system prompt has a pre-written creative ad shot language template for cat food. The user prompt simply references the brand variable. The shortcut for referencing variables is pressing Shift + { in English input mode.

Seedance vs. Happy Horse Video Generation Plugins

Seedance 2.0 Plugin

Seedance is ByteDance's AI video generation model, built on a Diffusion Model architecture. It supports both Text-to-Video and Image-to-Video modes. Version 2.0 brings significant improvements in motion consistency, image quality, and physical realism. The diffusion model works by starting from random noise and progressively denoising to generate the target content, which enables coherent and high-quality visual transitions in video generation.

You can find it in the Coze plugin library by sorting by "Recently Published." It comes in a fast version and a slow version. The slow version is recommended — it consumes only half the compute of the fast version, generates a 15-second video in about 5–10 minutes, and is actually faster than the web interface. It supports the all-purpose reference mode (default), with configuration options for API Key, prompt, duration, aspect ratio, and reference image.

Happy Horse Plugin

The configuration is essentially the same as Seedance, referencing the same variables. Both consume similar amounts of compute, and each has its own strengths — it's best to test and compare them in practice.

Reference Image Handling Tips

The plugin requires reference images in array format, but the start node passes in a single image. The solution is to add a "Simple Assistant" → "String to List" function before the video generation node to convert the single image into an array format.

Loop Querying and Video Result Output

Task Query Mechanism

After video generation, a task ID is returned. A loop query node checks whether the task is complete:

1. Set the loop type to infinite loop
2. Add the corresponding plugin's task query node inside the loop body (note: Seedance and Happy Horse use different query nodes)
3. Add a selector to check whether the returned URL is empty
4. If no result is found, use a timer to wait 60 seconds before retrying to avoid triggering rate limits
5. Once a result is found, exit via a break loop node

This polling mechanism is a common pattern for asynchronous task processing. Since video generation is computationally intensive, the server handles it asynchronously — it returns a task ID first, and the client periodically queries the status until completion. A 60-second wait interval strikes a good balance between getting timely results and avoiding API rate limiting from overly frequent requests.

Deduplicating Results

Since multiple queries may produce multiple outputs (some of which are empty), you can use the "Remove Empty Values from Array" function in the "Common Code Blocks" to ensure only valid video links are included in the final output.

Practical Tips and Notes

• During testing, use 720p resolution to save compute
• Before building, use notes to clearly document the logic of each node — it's a good habit to develop
• If you don't need loop querying, you can simply use the task ID to check results on the web interface's task query page

Key Takeaways

• The workflow core consists of just two nodes: LLM prompt generation + video generation plugin
• Seedance 2.0 slow version offers better value, consuming half the compute of the fast version
• Reference images must be converted to array format using a string-to-list tool before being passed to the plugin
• Loop querying should include a 60-second timer to avoid triggering rate limits
• Seedance and Happy Horse each have their own strengths, with similar compute consumption