Originally published at openvideos.ai. My vertical AI clips used to look muddy even when the photo was good. The model was not the problem. The crop was: I was feeding a landscape photo into a 9:16 job and hoping the tool would re-frame it....
Originally published at openvideos.ai.
My vertical AI clips used to look muddy even when the photo was good. The model was not the problem. The crop was: I was feeding a landscape photo into a 9:16 job and hoping the tool would re-frame it.
Here is the workflow I now run before I spend a single video credit.
The crop test
I reject a still for image-to-video if any of these are true:
The face is unreadable at phone size.
Hands dominate the frame.
There is readable signage behind the subject.
Two subjects compete for hero.
It is an extreme wide with a tiny face.
For a vertical clip I also check the thirds: face in the upper third, a quiet lower band for captions, and negative space in the direction of the camera move. I crop to 9:16 in my editor before upload, not after.
Then the prompt only carries motion
The still already contains the wardrobe, the location, and the light. Re-describing them gives the model two competing briefs. A motion-only prompt works better:
Keep the same woman, cream knit sweater, and golden café light from the start frame.
Compose for a phone: face in the upper third, quiet lower band for captions.
Camera: slow push-in over 5 seconds, stabilized.
Subject: she lifts the coffee cup slightly, holds eye contact.
No text, no watermark, no logos, no morphing hands.
One camera move, one subject action, one vertical placement instruction.
What it costs (registry numbers, re-check in Create)
Model id
Settings
Credits
seedance-v1-5-pro-i2v-720p
5s, 9:16, audio off
75
seedance-v1-5-pro-i2v-720p
5s, 9:16, audio on
150
hailuo-02-i2v
768p
200
hailuo-02-i2v
1080p
300
wan-2-5-i2v
5s, 480p
140
wan-2-5-i2v
5s, 720p
260
One honest note: seedance-2-i2v is registry-tipped for anime and AI characters only, so for a real-person photo I default to Seedance v1.5 Pro at 5 seconds, 9:16, audio off (75 credits).
Checkpoint before take one
Still looks postable at 9:16 on its own.
Category is start-frame-video, not text-to-video.
The prompt names one camera move, one action, and the vertical placement.
The Create total matches my budget.
The full workflow with a failure table and a copy-ready prompt lives on the original post:
→ Turn a Photo into a 9:16 AI Short