We run JME Freestyle AI, a small tool that takes a single uploaded photo and returns a 15-second vertical rap video. The interesting part is not the model - it is that we ask the user to choose a template instead of writing a prompt.
Why templates instead of prompts
Prompt-based video generation pushes the failure mode onto the least prepared person in the loop. If someone has never written a shot list, "cinematic dolly-in, 35mm, moody lighting" is a coin flip. A template encodes the shot list once, and the user only supplies what they actually have: one photo.
Each template fixes three things:
- the music style and tempo, so the cut points are known in advance
- the camera choreography (push in, whip pan, static), so the motion is plausible for the source image
- the on-screen text structure, so the lyrics land on beats rather than drifting
That converts an open-ended generation task into a bounded one. Bounded tasks are cheaper to render and far easier to evaluate.
Rendering policy that actually matters: shortest duration
We render at 720p vertical and apply a shortest-duration policy across every pass, including one-time passes. In practice: if the template says 15 seconds and the useful content finishes sooner, we trim. Nobody pays for dead air, and the clip is immediately usable on TikTok, Reels and Shorts without a re-export.
Trimming also hides a class of model artifacts. Long generations accumulate drift - faces soften, backgrounds swim. Cutting at the last coherent frame is a cheaper fix than a quality filter.
Model layer
Generation runs on Wan 3.0 through an API partner. We deliberately do not train a custom face model: the product promise is "your photo, one minute, no setup", and a hosted image-to-video model plus strong templates gets us there. The trade-off is that we inherit the provider's failure modes, so the template library is where our own quality control lives.
Safety and rights
Two rules are non-negotiable and both are product decisions, not legal boilerplate:
- You upload images you are allowed to use. We do not scrape faces.
- Safety checks run before generation, not after. A blocked render costs the user nothing.
The templates are original arrangements. We are not affiliated with any artist or label, and nothing in the product claims otherwise.
What we would tell another team building this
- Constrain the input. One photo is a better contract than "describe your video".
- Put the taste in the template, not in the prompt.
- Bill by the finished clip, not by the GPU minute.
- Ship the trim logic before the style transfer.
If you want to see the template list, the tool is live at jmeukrap.com, and credit packs and one-time passes are on the pricing page.
Top comments (0)