We recently produced and delivered a promotional video made with AI — 3 minutes 30 seconds, 1080p. Under the contract we cannot show it as part of our portfolio, so instead of the result we are recording the process: the order we worked in, what went as planned and what did not, and what it actually cost.
It took 17 working days.
Script and storyboard
The script was written with Claude and the storyboard drawn with ChatGPT — settle the flow in words first, then turn scenes into pictures.
The storyboard then served as the reference for two later stages: it became the shot list for the model shoot, and the brief for each scene during generation.
We shot a real model
An AI video does not mean no one was photographed. We signed a contract with a real model and ran a photo shoot.
The contract limited use of the photos to AI video production, and usage rights to the resulting video only. Since the model's face becomes the reference for the generated footage, we felt the scope of use had to be clear from the start.
The shoot was broken down closely against the storyboard: pose, expression and framing were set per scene, for around 30 photos across multiple cuts. That preparation made the biggest difference later. With a reference image for every scene, generation needed few retries — most scenes came out as intended in one or two attempts.
Generation — Claude connected to Higgsfield MCP
We fed the model shots, script and storyboard to Claude and generated the video through Higgsfield MCP, using Seedance 2.5 at 1080p.
The video is built from many short scenes joined together. Cut transitions and the connecting shots between scenes were designed separately so the full 3 minutes 30 seconds flows without breaks.
Sound: played by hand, imitated by voice
The background music was made in Suno — not from a prompt alone, but from guitar and piano we recorded ourselves, turned into an instrumental track.
Sound effects were made in ACE Studio: we imitated each sound with our voice, then generated effects that follow that shape and placed them in the edit.
Finishing the traditional way
Final editing and correction were done the traditional way in Adobe Premiere Pro.
For the logo ending, we built a 3D logo animation with Claude connected to Blender MCP, then brought it into After Effects to finish.
Cost and fee
Approximate costs, all excluding VAT:
- Model fee — KRW 700,000
- AI usage — KRW 1,250,000 (usage converted to token cost)
- Planning and project management — KRW 1,000,000
- Production labour — KRW 1,000,000
- Total — about KRW 4,000,000
The project fee was KRW 5,500,000. It took 17 working days.
What did not go as expected
Generation itself was easier than expected. Because the reference shoot was so detailed, most scenes came out in one or two attempts. The cost and time went elsewhere.
Revision requests consumed most of the tokens. When the client asked for a change after seeing a generated scene, regenerating that scene alone was often not enough — its connection to the scenes before and after changed, so the adjoining clips had to be regenerated too. Most of the AI spend went to this revision work, not the first pass.
Some transitions would not connect through generation. In a few places the join between clips did not flow naturally, and we re-edited those sections the traditional way.
If we did it again
We had considered generating at 720p first for client approval, then regenerating approved scenes at higher resolution — a way to cut costs during the revision-heavy early stage.
But we found that every regeneration comes out slightly different. A scene approved at 720p is not guaranteed to come back identical when regenerated, so we concluded this approach does not work with current video generation technology.
Revisions, then, are a cost that is hard to reduce. When fixing one scene means regenerating its neighbours, agreeing the scope and number of revisions at the quoting stage is how the budget holds.
Closing
On this project, AI handled the script and storyboard, video generation, music and sound effects, and the logo animation. The model shoot and contract, final edit and finishing were done by people. Generation finishing in one or two tries, and the broken joins being repaired, both came down to that work at either end. AI changes how a video is made; it does not lighten the weight of preparation and finishing.
If you are considering an AI-produced promotional video, talk to us.