Every student media club gets the same requests: a 10-second promo for the robotics open night, a vertical teaser for the film society, a short recap clip a TA can drop into the course page after a tough lecture. So I tested whether an AI video model could take on that first draft, and measured what came back: whether the words on screen are readable, whether simple explainers follow the prompt, how long a clip takes and what each usable one costs.
What I tested and how
I used Kling 4.0 Flash, model ID kling-v4-flash. It makes 720p clips of 3 to 20 seconds, from a text prompt or from one first-frame image, and every clip comes with a generated soundtrack. I submitted 17 jobs. Two were rejected on purpose before they started: asking for 1080p returns "only supports resolution 720p", and sending an aspect ratio with an image returns an error saying image-to-video follows the first frame's ratio. Neither was charged. The other 15 all completed on the first try.
A request is one JSON body, then you poll the task:
POST {API_BASE}/v1/videos/generations
Authorization: Bearer YOUR_KEY
Content-Type: application/json
{
"model": "kling-v4-flash",
"prompt": "Flat 2D motion graphics on a soft green background: a stack of three cartoon books bounces in, then the text STUDY JAM appears above it and FRI 6 PM below it, bold dark letters, centred, fully inside the frame. Light playful music.",
"duration": 3,
"aspect_ratio": "16:9"
}
GET {API_BASE}/v1/tasks/{task_id}
For image-to-video, drop aspect_ratio and add "image_urls": ["https://.../poster.png"]. For each text clip I pulled 25 evenly spaced frames and ran OCR on them (6 frames for the other clips), so text results below are counted, not eyeballed.
Test 1: Event title cards
Three short cards, each run more than once: CAMPUS FILM NIGHT, STUDY JAM with FRI 6 PM, and LIBRARY OPEN 24/7 DURING FINALS. Each 3-second clip cost $0.26.
Spelling was perfect: OCR found no misspelled word in any of the eight clips, including "24/7" and "6 PM". The one failure was layout. My first prompt did not mention margins, and the title grew until it ran off both edges, ending as "MPUS FILM NIG".
After I added "centred with wide margins, fully inside the frame", all six later 3-second runs fit. Text needs time to appear, though: the full line was first readable after 0.4 s for the film card, 0.9 to 1.1 s for STUDY JAM and 1.0 to 1.7 s for the library card. In a 3-second clip that leaves under two seconds of readable text. One library run set the words at about half the size of the other, small on a phone. For a story slot, I'd generate 5 seconds and trim.
Test 2: Lecture-recap explainers
Next, two 5-second explainers ($0.43 each) of the kind a TA might want. Biology: a leaf, sun rays from the top left, blue CO2 bubbles flowing in, red O2 bubbles flowing out, a water droplet rising into the stem. Economics: a supply line shifting right on a price-quantity graph.
The photosynthesis clip got all five elements in the right places and directions, plus a slow zoom I did not ask for. The economics clip missed the one thing that mattered. The supply line stayed where it was, and instead a black dot slid down the demand line, which teaches the wrong idea. There were no axes either.
Test 3: Animating a poster and a slide you already made
Most clubs already have a poster, so I animated one I had designed, plus a simple lecture slide.
This was the most dependable result in the whole test. In both poster runs, OCR returned the exact poster text, all four lines down to "THURSDAY 7 PM · ROOM 214", in every sampled frame. With the "keep the layout" prompt only 2.2% of pixels changed between first and last frame, all of them in the decorative circles.
The slide failed twice. I asked for the red supply line to shift right while everything else stayed fixed. In run 1 nothing moved: 0.3% of pixels changed in five seconds. In run 2, with a more explicit prompt, both lines thickened as if highlighted, but neither moved.
So it preserves your design and adds ambient motion, but it won't perform a specific change to a diagram. Animate graphs in your slide software.
Time and cost
Length barely changed the wait. The 3-second cards took a median of 59 seconds, 5-second clips about 71 seconds, and the 10-second clip 70 seconds. Billing was per second, $0.0857 for both text-to-video and image-to-video, which matches the 720p rate listed on the Kling 4.0 Flash API page when I tested. The whole set of 15 clips cost $5.66. Eleven of them were usable as they came out, so I paid about $0.51 per usable clip, failures included. Text cards on their own came to about $0.38 per usable clip.
Test 4: A vertical promo with people
For social, I made one 9:16 clip: three students carrying a cardboard robot across a campus lawn. It came back as 720x1280, sharp and convincing, with natural walking and believable faces. That realism is exactly why it needs a label, which I come back to below.
Disclosure: how we label AI video in club and class use
None of the 15 files carried C2PA Content Credentials, and the metadata held only a generic encoder tag. Every clip also had a stereo AI soundtrack. Nothing in the file tells a viewer it was generated, so the label has to come from you. Our rules:
- A visible "AI-generated video" tag on screen for the whole clip when it shows people or realistic places, plus a line in the caption.
- No generated versions of real students, staff or professors, and no real campus buildings, logos or mascots.
- For course material, follow your instructor's or department's AI policy and say in the post which visuals are generated.
- Dates, times and room numbers are checked by a second person before posting, even when the OCR reads them correctly.
- Replace or check the generated audio. Music that sounds like a known song is still a problem.
Bottom line
For a student media team, this is a solid first-draft tool for title cards and poster animations: spelling was right in every clip, existing designs survived untouched, and a clip costs well under a dollar. Plan for a minute's wait per clip, ask for margins, and give text time to appear. Keep it away from explainers where the motion itself is the lesson, like a curve shifting, and label anything that looks real.