What is CinePrompt and why does it exist?
A cinematographer's toolkit collapsed into a text box overnight. CinePrompt is what happened next.
Sora died yesterday
OpenAI shut down Sora on March 24. The app, the API, the Disney deal, the whole thing. One of seven CinePrompt models just went dark. The vocabulary did not.
The actor is the reference image
Val Kilmer died last year. A generative AI built from his recorded existence will star in the film he never got to make. The tools are the same ones in your browser tab right now.
The viewfinder came back
Runway demonstrated real-time AI video generation at GTC. Under 100 milliseconds to first frame. The delay between prompt and output was where the craft lived.
They yassified the ground truth
NVIDIA fed generative AI complete structured game data and the AI still decided everyone needed to be prettier. Gamers recognized the beauty bias in thirty seconds.
Everybody picked the pretty one
Grok Imagine generated 1.245 billion videos in January. Arena leaderboards crowned it number one. The judges were not filmmakers and the test was not filmmaking.
Free was never the product
xAI locked free users out of Grok Imagine overnight. The people who lost the most are the ones who never built anything portable.
The ground truth
NVIDIA just proved that structured data makes AI output controllable. Filmmakers figured this out with a smaller budget.
Real footage needs proof now
A head of state posted an authentic video. An AI flagged it as a deepfake. Millions believed the machine. This is where the quality gains land.
Runway left the set
Runway launched a world simulation engine and started hosting competitors' models. The most filmmaker-focused platform is becoming a platform for everyone.
Spielberg does not need a text box
Steven Spielberg told SXSW he has never used AI in any of his films. The crowd cheered. They were cheering for the wrong thing.
What the model watched
ByteDance suspended Seedance 2.0's global launch over copyright disputes. The real story is what every model's training data means for what you can prompt.
You cannot prompt ugly
AI video models have a systematic beauty bias. Ugliness is one of cinema's sharpest tools and the hardest thing to prompt for.
The model moved in
NVIDIA just made AI video generation run on the box under your desk. The cloud was a landlord. Local is ownership. Both have consequences.
The chatbot ate the camera
OpenAI is putting Sora inside ChatGPT. Video generation just moved from the studio to the group chat.
The cut is still yours
Nineteen articles about what AI video models can generate. Zero about the decision that turns footage into a film.
Good enough to fool someone
X started penalizing AI-generated war videos without disclosure labels. The reason says more about the craft than any benchmark.
Nobody told the model where to put anything
Seventeen articles about what to generate. Not one about where it goes in the frame. Composition is the most fundamental visual decision and the one most surrendered to defaults.
Netflix just bought the gap
Netflix acquired a tool that teaches AI to understand filmmakers. That gap between knowledge and comprehension is not a bug. It is the product category.
The generate button is a commodity
Every AI video platform sells the same API call in a different wrapper. CinePrompt ships the knowledge layer and lets you bring your own keys.
You forgot to build the room
Fourteen articles on cameras, lights, actors, color, and sound. Not one mentioned the walls. The environment is an entire department with no analog in a text prompt.
The face works. The performance doesn't.
Thirteen articles on cameras, lenses, lights, color, and sound. Zero on the person standing in front of all of it. Directing an AI actor is the hardest prompt you will ever write.
Time is the one dial you cannot turn
Twelve articles about controlling the image. None about controlling when things happen inside it. Time is a filmmaker's sharpest tool and a prompter's locked door.
Stop asking which model is best
Every week another comparison article ranks five AI video models on a single axis. The question itself is the problem.
Frame to Motion is the whole point
The img2vid workflow that treats a reference image as half your prompt. Two generations, each with a focused job, producing a result neither could achieve alone.
Nobody is prompting for sound
Four of five AI video models generate native audio now. Most people are still prompting like sound does not exist.
The prompt stopped being the product
Eight articles about what to type. This one is about everything that happens around it. The text box is shrinking. Not in size. In relative importance.
The anatomy of a prompt that actually works
You know which words to use. This is about what order to put them in, how many to use, and why most prompts are architectural disasters.
How to make five AI clips that don't look like five strangers
Every AI video model can make a beautiful shot. Almost none of them can make two that belong together. Here is how to fight for sequence coherence.
Lighting words that actually shape the frame
"Dramatic lighting" is the new "cinematic." Here is what to type when you want the model to actually move the shadows.
Lens specs that do something vs lens specs that don't
Focal lengths, f-stops, lens names, and why the most technical-sounding part of your prompt is probably doing the least work.
Color words that do something vs color words that don't
Film stock names, color grading terms, and the gap between what you type and what the model sees.
Camera movement keywords that actually work
Dolly, crane, tracking, orbit. Some of these words move the camera. Some of them do nothing. Here is what each model actually hears.
Why "cinematic" is the most useless prompt word
Every AI video model already thinks it is being cinematic. Saying the word does nothing. Here is what to say instead.