AI UGC content system with AI agents (n8n template)

The Automated UGC Ad Architecture

0:00 to 4:29 in the video, about 5 minutes

Modern video generation models allow marketers to create user-generated content without filming physical actors or shipping product samples. The core system connects a Google Sheet to an automation platform called n8n, routing product records to different AI video generation pipelines based on user selection.

The workflow monitors the spreadsheet for rows marked with a specific status, ensuring it only processes ungenerated items one at a time. A routing switch node evaluates the chosen video model and directs the data down one of three branches: direct Google Veo 3.1, direct OpenAI Sora 2, or a hybrid pipeline using Nano Banana image generation paired with Veo 3.1.

• Google Sheets acts as the lightweight database for product photos, customer profiles, and scene descriptions.
• A Switch node dynamically routes rows to specialized model paths.
• HTTP Request nodes communicate with model endpoints through the Kie AI API gateway.

Comparing Direct Video and Hybrid Image-to-Video Pipelines

4:30 to 5:46 in the video, about 1 minute

Feeding a raw e-commerce product packshot directly into video generation models often yields visual awkwardness. In testing creatine gummies, hair shine spray, and portable neck fans, both Sora 2 and direct Veo 3.1 tended to show the product image frozen as the video's opening frame before motion began.

To solve this issue, the hybrid pipeline introduces an intermediate step. It uses an image generation model called Nano Banana to transform the plain product image into a realistic still photo of a person wearing or holding the product in an appropriate setting, which then animates seamlessly via Veo 3.1.

• Direct image-to-video often leaves the initial still product packshot visible on the first frame.
• Sora 2 restricts users from inputting realistic human faces into its image-to-video endpoint due to content guardrails.
• Veo 3.1 accepts realistic AI-generated human images, allowing it to animate pre-staged actor stills.

Crafting Prompts with Dedicated AI Agents

5:47 to 14:28 in the video, about 9 minutes

The workflow employs large language model agents running GPT-4o-mini to convert structured product data into descriptive prompts tailored for image and video synthesis.

The image prompt agent takes the product title and setting description, applying system prompt rules that enforce handheld selfie aesthetics, natural outdoor lighting, authentic skin textures, and strict product visual fidelity without altering logos. A vision-enabled model then inspects the resulting generated still image to catalog specific visual details like clothing colors, surroundings, and hand positioning.

• System prompts enforce vertical selfie-style framing, realistic facial imperfections, and proper lighting.
• A vision node analyzes the generated image to supply context back into the downstream video prompt.
• Regular expression cleaning removes newline characters and smart curly quotes that could break JSON API payloads.

Managing Asynchronous Generation and Sheet Updates

14:29 to 22:38 in the video, about 8 minutes

Image and video generation APIs operate asynchronously, requiring the workflow to manage tasks that take tens of seconds to render. After dispatching generation tasks to the API gateway, the workflow enters a polling loop using Wait nodes and conditional logic.

The system waits in short increments, queries the task status using the generated task identifier, and evaluates whether the state equals success. Once the rendering is marked complete, n8n extracts the video file URL, updates the original Google Sheet row status to finished, and records the final link.

• Wait nodes pause execution in intervals ranging from 5 to 10 seconds between API status checks.
• If-nodes verify the job response status, looping back to wait nodes until the file is ready.
• Google Sheets nodes match on row index numbers to update statuses and insert finished video URLs.

Quality Differences and Cost Trade-offs

22:39 to 24:17 in the video, about 2 minutes

Evaluating the generated video outputs reveals distinct stylistic differences across the tested platforms. Direct Veo 3.1 generations occasionally suffered from heavy HDR styling, unnatural skin contrast, and minor hallucinations, such as turning a gummy jar into a pouch.

Sora 2 exhibited strong motion and natural camera movement, though its inability to directly accept staged actor stills restricts its use for purely automated selfie video ads. Financially, generating a 10-second Sora 2 video costs about 15 cents, while the hybrid Nano Banana image (2 cents) plus an 8-second Veo 3.1 Fast video (30 cents) totaled 32 cents per creative piece.

• Option 1 (Nano Banana + Veo 3.1 Fast): Approximately 32 cents per finished video.
• Option 2 (Direct Sora 2): Approximately 15 cents per finished video.
• Sora 2 offers high cost-efficiency for bulk testing, while the hybrid pipeline provides superior visual coherence for automated publishing.

Final Thoughts

Combining automated orchestration tools like n8n with specialized AI agents bridges the gap between raw generative models and viable commercial assets. While standalone video generators continue to struggle with static opening frames and content safety limitations regarding human faces, a multi-stage approach pairing image staging with video animation yields far more authentic UGC results.

As video models advance toward next-generation iterations, costs will likely drop while generation fidelity and strict product adherence improve, making fully automated marketing creative increasingly viable for e-commerce brands.