Krea releases five free realistic video prompts

METAL LAB downloaded that prompt pack directly and has translated the full text of all five prompts into Korean below [translated here into English for this article]. There's no need to download the file separately — readers can simply pick what they need from this article and copy it.

Who made Seedance 2.5

One thing worth clarifying here: Krea is not the company that developed these videos. Krea is a hosting platform that brings together various generative models so they can be run on the web, and the engine that actually produced the videos — Seedance 2.5 — is reportedly a video generation model series developed by ByteDance Seed. It works by converting text or image input into short video clips, and in recent months it has drawn attention for output that's difficult to distinguish from real footage. Krea's post here is less a claim of "we built this model" and more a case study: "here's what this model produced on our platform."

The shared skeleton across all five prompts

Lay the five prompts side by side and the structure is identical — only the scene changes. What creates the realistic texture isn't an adjective like "realistically," but sentences that deliberately spell out camera flaws. Shake, water and dust on the lens, autofocus hunting for focus, distorted mic audio — these instruct the model to produce footage that looks carelessly shot on location, not a polished take.

Component	Shared instruction across all five
Length/cuts	10 seconds · one shot · no cuts · 16:9
Camera gear	Phone, action cam, waterproof housing, chest mount — all amateur handheld gear
Camera flaws	Shaky/wobble/heavy shake, rolling shutter judder, autofocus hunting, motion blur
Lens grime	Bubbles passing over the lens, water running down it, splattered dust, lens flare
Lighting	All daytime · natural light only
Audio	On-board mic · heavily distorted sound · ambient sound only
Banned phrases	Photoreal, physically accurate / no CGI look, no slow motion, no music, no text, no logos

The bottom row's banned phrases appear word-for-word identical across all five prompts. Stripping out the slow motion and smooth CG tone that video generation models default to is the core mechanism of this prompt pack.

① Downhill skateboarding

Format — 10-second handheld phone video, one shot, no cuts, 16:9

Scene — Daytime. A steep city hill lined with parked cars on both sides, a busy intersection visible far below.

Camera — Shot from the passenger window of a car struggling to keep pace with the speed. The frame shakes, wind hits the mic, rolling shutter judders, autofocus hunts for focus.

Action — A skater bombs the hill in a full tuck. The board wobbles as he weaves between two moving cars, arms out for balance. The filming vehicle falls behind, then catches back up. As the intersection nears, he carves hard across the full width of the road to shed speed, then rises out of the crouch to clear it.

Audio — Phone mic, heavily distorted. The roar of urethane wheels rolling over asphalt, wind slamming the mic, the filming car's engine, shouting and laughter bursting from inside the car.

Banned — Photoreal, physically accurate. No CGI look, no slow motion, no music, no text, no logos.

Phrases to keep in English — 10-second handheld phone video, one shot, no cuts, 16:9 · bombs the hill · full tuck · rolling-shutter wobble · autofocus hunting · on-phone mic, heavily overdriven

② Surfer wave

Format — 10-second waterproof housing action cam video, one shot, no cuts, 16:9

Scene — Daytime, open ocean. Heavy grey swells and spray hanging suspended in the air.

Camera — A waterproof housing held by a swimmer positioned in the impact zone where waves break. The frame splits half water, half air, water streams down the dome port, the horizon tilts wildly.

Action — A massive wave rises and begins breaking straight toward the camera. A surfer drops in from the crest and races along the face, spray blasting off the lip above. The wave keeps growing over the camera until the lip pitches forward and finally collapses entirely over the lens. The frame is swallowed in a cloudy churn of water and bubbles.

Audio — On-camera mic. Wind, a swelling low roar of water, the impact of the lip crashing down, then a sudden muffled churn.

Banned — Photoreal, physically accurate. No CGI look, no slow motion, no music, no text, no logos.

Phrases to keep in English — water-housing action camera · half water half air · dome port, water sheeting · horizon tilting wildly · drops in, races along the face · muffled churn

③ Galloping horses

Format — 10-second chest-mounted action cam video, one shot, no cuts, 16:9

Scene — Daytime, dense forest. Low sun drives hard shafts of light between the trees.

Camera — Mounted on the chest of a rider galloping at full speed. Wide-angle lens, shaking hard in sync with the horse's stride, motion blur, lens flare strobing every time it passes a tree trunk.

Action — Three horses gallop at full speed just ahead. Dirt and leaves kick up from the hooves, riders lean low over the horses' necks. The camera closes in on them, weaving between tree trunks, low branches whipping past close to the lens. The ground streams by below. One horse pulls ahead through a gap between two trees.

Audio — On-camera mic. Roaring wind over it, hooves hammering the dirt in rhythm, the horses' ragged breathing, snapping twigs, a rider shouting.

Banned — Photoreal, physically accurate. No CGI look, no slow motion, no music, no text, no logos.

Phrases to keep in English — chest-mounted action camera · full gallop · hard shafts of low sun through trees · lens flare strobing · hooves hammering the dirt

④ Deep sea beneath the ice

Format — 10-second underwater action cam video, one shot, no cuts, 16:9

Scene — Daytime, beneath a frozen lake. The ice ceiling overhead is a pale blue-white, riddled with cracks, sunlight falling through in soft columns. The water is glass-clear, the depths below nearly black.

Camera — Handheld underwater camera with a wide-angle lens. A faint shake, small bubbles drifting past the lens, color shifting toward cyan with depth.

Action — A freediver in a black wetsuit and monofin, no tank, no bubbles, swims past the camera just beneath the ice, one hand trailing along the ice overhead. The camera follows, slowly descending. The diver rolls onto her back to look up at the ice above, then spots a bright hole ahead and swims toward it.

Audio — Muffled underwater ambience, the ice groaning and cracking, a low hiss of water against the camera housing. No breathing sounds included.

Banned — Photoreal, physically accurate. No CGI look, no slow motion, no music, no text, no logos.

Phrases to keep in English — underwater action camera · pale blue-white ice ceiling · sunlight in soft columns · color shifting toward cyan · freediver, monofin, no tank, no bubbles · ice groaning and cracking

⑤ Off-road trail

Format — 10-second action cam video, one shot, no cuts, 16:9

Scene — Daytime, dirt motocross track. Churned brown dirt and tire tracks, sparse trees behind.

Camera — Sitting low at the landing of a big tabletop jump, facing back toward the takeoff. Wide-angle lens, faint shake, dust sitting on the lens, sunlight leaking into the frame.

Action — A rider clears the jump, throwing the bike sideways into a whip. He corrects mid-air and lands hard on the landing ramp, suspension bottoming out. He passes directly over the camera, roosting dirt onto the lens.

Audio — On-camera mic, heavily distorted. A two-stroke engine screaming off the jump then cutting off, wind, the crunch of landing, dirt hitting the housing, chain and suspension noise.

Banned — Photoreal, physically accurate. No CGI look, no slow motion, no music, no text, no logos.

Phrases to keep in English — sitting low at the landing of a big tabletop jump · whip, throwing the bike sideways · suspension bottoming out · roosting dirt onto the lens · two-stroke engine screaming

How to replicate this yourself

1. Copy one of the five prompts above as a whole paragraph. Keep the order — Format → Scene → Camera → Action → Audio → Banned — intact. That order is itself the prompt's structure.
2. Select Seedance 2.5 on the Krea platform and paste it in. Entering it in another language will still capture the scene, but mixing in the English phrasing noted above for camera and lens instructions produces more accurate results.
3. If the output comes out smooth like CG, first check whether the banned phrases at the end (no CGI look, no slow motion) survived intact. The next thing to adjust is the camera-gear line — swapping "phone," "action cam," or "chest mount" changes the entire texture.
4. When adapting to a different subject, swap out only the Scene paragraph and leave the other five lines as they are. The shared skeleton in the table above is what produces the realistic texture, so leaving it untouched reproduces the same tone even with a different subject.

The original source doesn't specify pricing tiers or sign-up requirements, so the scope of free usage should be confirmed directly on the platform.

Already-proven use cases

This isn't the first time Seedance 2.5 has been used for content production. Higgsfield released an animated opening sequence made with the model as a YouTube Shorts video on August 10, and two days later, on August 12, released all the prompts and assets used to produce a new episode of its K-pop action series, "Zephyr: Special." At the time, Higgsfield described having "pushed the model to its limits," and opened up access so that participants in its own $1 million film festival could replicate the work directly. The fact that two separate platforms have released outputs from the same model in quick succession shows that Seedance 2.5 is being adopted across multiple services.

Why companies are releasing prompts

There's a notable trend of companies working with video generation models releasing not just outputs but the prompts themselves. Just as Higgsfield opened up its assets entirely for film festival participants, Krea too is sharing prompts that produced good results rather than keeping them hidden. This reads as a strategy that both demonstrates the model's capabilities and keeps users on the platform by encouraging them to replicate the results themselves. It also means the barrier to producing realistic-looking video is being lowered to the level of a single prompt paragraph.

So what changes

The criteria for choosing a video generation model are shifting from "which company built it" to "which platform makes it easiest to produce good results." At a moment when platforms like Krea are racing to publish prompts and usage tips for the same underlying model, users now have more options to compare outputs from the same Seedance 2.5 model across different platforms. And the starting point for that comparison is a single paragraph of prompt text, like the ones translated above.