Overview

L’Artisan Parfumeur is my first UGC advertisement created through an AI-assisted workflow. During the process, I quickly realized that UGC advertising differs significantly from other types of commercials or short-form films. In this format, character consistency and believable mouth sync are especially critical—often even more important than composition, camera transitions, or other purely visual considerations. With that in mind, I set out to create a realistic fragrance UGC ad featuring a woman in her thirties speaking directly to the camera, intercut with B-roll footage and ending on a product close-up. My goal was to make the final video feel as close as possible to authentic mobile-shot user-generated content while still maintaining a polished luxury aesthetic.
 

To begin, I developed the concept, script, and shot structure through conversations with large language models. I used this process to quickly shape the direction of the ad, defining both the speaking portions and the supporting B-roll moments. I then used Nano Banana Pro to generate the character design and environment design. After that, I collected real product photography of this L’Artisan Parfumeur fragrance from online sources and used Nano Banana Pro to generate additional product images from multiple angles. With the help of large language models, I was able to refine the script and storyboard efficiently and establish a clear visual and narrative direction before moving into video generation.
 

I first attempted to generate the video using Seedance 2.0. However, at its current stage, Seedance has fairly strict limitations when it comes to using character references consistently, which made the video generation unsuccessful for this project. I then turned to Kling 3.0 Omni and ran several rounds of testing. Through that process, I found that Kling was not well suited for this particular type of UGC advertisement—especially for shots where the character stands in place and delivers lines directly to camera. Those shots often felt stiff and unnatural. When generating a full 12-second video, the internal edit timing also tended to feel awkward, while generating the video shot by shot created major consistency issues with the character. After evaluating those results, I decided to move away from Kling and instead use Sora 2 Pro, which currently produces footage that feels much closer to real smartphone-shot video, something that is essential to the visual language of UGC advertising.
 

Because I had already prepared start frames for different shots in Nano Banana Pro, the video generation process in Sora 2 Pro went relatively smoothly. Once I had generated and organized all usable clips, I moved into Premiere Pro for editing, music, and color correction. I then used ElevenLabs to generate an appropriate voiceover for the final piece. After the edit was locked, I upscaled the video in Topaz Video, and finally used CapCut to manually adjust and refine the subtitles before completing the advertisement.

This project gave me a much clearer understanding of both the strengths and limitations of current AI video workflows in the context of UGC advertising. Compared with more cinematic or stylized projects, UGC demands a stronger sense of naturalism, especially in facial performance, mouth movement, and the subtle realism of someone speaking casually to a camera. In that sense, this project was less about creating visually dramatic shots and more about finding the workflow that could best preserve authenticity. While current tools still have clear limitations, they are already highly effective for rapidly developing scripts, storyboards, style frames, product angles, and early visual concepts. For me, this project was an important step in understanding how AI tools can be used not only for stylized image-making, but also for commercially viable, realism-driven short-form advertising.