DOUBLE TROUBLE Somewhere between directing a short film and arguing with an AI, I learned that mo...DOUBLE TROUBLE Somewhere between directing a short film and arguing with an AI, I learned that mo...
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
DOUBLE TROUBLE
Somewhere between directing a short film and arguing with an AI, I learned that more control does not always mean more control.
For the CapCut × Contra Seedance 2.5 30-Second Cinematic Video Challenge, I wanted to make a one-take thriller built around a simple idea: an apparently harmless old man walks into an active bank robbery, keeps moving like nothing is happening, and quietly turns out to be much more dangerous than he looks.
The first version was heavily engineered. Character sheets, keyframes, storyboard, detailed choreography, reference images, long prompt. It sounded disciplined, but the result was the opposite. The references started competing with each other, transitions became less natural, and the pacing drifted.
So I stripped it back.
I removed the storyboard and keyframes, kept only the references that were actually doing useful work, and made the video prompt much more focused. The result immediately felt freer and more cinematic. It still had imperfections, but it also had something the heavily controlled version didn't: room for Seedance to interpret the scene.
That became the real experiment.
For the final attempt, I kept the streamlined reference setup but focused on one thing: making the story progression clear without trying to choreograph every microscopic movement. The robbery, the three escalating encounters, the punch, the physical prosthetic reveal, and the final beat all needed to read as one continuous piece of storytelling.
The biggest lesson was simple:
Less is not necessarily less control. Sometimes it is better control.
A prompt can describe everything and still make the model lose the story. The better approach was deciding which details actually mattered, locking those down, and leaving enough room for the model to make the sequence feel alive.
The final film was generated with Seedance 2.5 in CapCut Video Studio. The reference images were created separately in ChatGPT, and Claude was used throughout the workflow for planning, critique, prompt iteration, and reconciliation.
I also made a separate workflow video explaining the process and what changed between the attempts. The biggest surprise was not the final image quality. It was realizing that the strongest version came from removing things rather than adding them.
📹Created with: CapCut Video Studio + Seedance 2.5 🎞️Workflow: ChatGPT for references, Claude for planning/critique, Seedance 2.5 for the final film
🪄Original prompt + workflow: included in the comments below.
#contraseedancechallenge #CapCutCPP #AIFilm
Ruffa's avatar
DOUBLE TROUBLE Prompt + Workflow Photorealistic live-action thriller. Inside a compact bank (Image 7). Exactly three robbers: Leader (Image 2, center-right), Playful One (Image 3, left), Money Guy (Image 4, right). An elderly man (Image 1) with a wooden cane (Image 5) and...
Back to feed
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started