AI video generation
Photo to video AI works like this: you upload a still, write two or three lines about the motion, and the model imagines what happens next. Your photo becomes the first frame of the clip. That is how a product shot turns into an ad, and a series of construction photos turns into a timelapse. I spent 13 years shooting and running a production company, and since 2023 AI has been part of how I work. My "AI Wedding and the Bride's Reaction" video reached 2.4 million views (my own site data as of November 5, 2026). Here's how I actually do it.
Short answer: 1) the photo is the first frame, and the AI generates the motion from there; 2) describe one or two movements and the camera, not the whole plot; 3) a product shot can be almost any quality, a face needs a sharp, well-lit photo; 4) there are two workflows: one photo plus a script (fast, less control) or your own storyboard of several frames (slower, but the clip comes out the way you planned it); 5) treat the first generation as a test, the usable one is usually the second or third; 6) speed the finished shot up slightly in the edit.
A video model doesn't "play" your photo like footage. It takes the image as frame one and invents every frame after it: how the camera moves, where the person walks, how the light falls. Two things follow from that.
First: the photo sets everything that's in the frame, and the prompt only sets the motion. You don't need to describe the face, the outfit or the color of the walls. They're already in the picture.
Second: you can't get other angles of the same place from a single photo. The model has never seen what's behind your subject, so it starts breaking physics. No generator can reliably do that today. If you need different angles, generate them from scratch, together in one request, instead of trying to "rotate" an existing photo.
And the main point: the animation isn't what matters most. The story is. A moving picture with no idea behind it gets scrolled past.
A lot depends on what's in the shot.
Products. For a product video, a product photo of almost any quality is enough. In my lesson on ad videos I deliberately used an old slipper to prove the point: you can make a commercial out of pretty much anything. I clean the photo up first in an image generator with a short request. Mine was: "Place the slipper on a clean white background for a product demo, remove everything else, and make the slipper look brand new." Then the video model gets a clean frame to work with.
People and faces. Here you need a sharp photo with clear light, and the rules change: one face in the frame, one action, a short prompt. I wrote a separate guide on this: how to animate a photo with AI without getting a creepy face.
Places, construction, interiors. For a timelapse, a series of photos works, not a single frame. When I tried to show a renovation from just two photos, I got a lot of generative garbage. The more in-between frames you give the model, the better the result.
In my lesson on ad videos I split the work into two approaches.
Semi-controlled: one photo plus a script. Upload the product photo, ask a text model to write a script, and hand that script to the video model. My request to the text model is simple: "Study the photo and write a shot-by-shot script for a short commercial for this product." Claude, ChatGPT or Gemini will all handle it; it's not a hard task. You get less control but more flexibility: the model will suggest moves you wouldn't have thought of.
Controlled: your own storyboard. You build the frames you want to see first, and the video model only brings them to life. When the clip has to come out exactly as planned, there's no way around it: create the image, then animate that image. That's how I made a phone ad from a single multi-panel storyboard image, and it turned into a complete commercial.
For first experiments and social posts, take the first. For a client where every frame matters, the second. The generator is a swappable part; the system stays the same.
I always aim for fewer actions. The less you ask for, the fewer chances the model has to get it wrong.
The template I use myself: who is in the frame, where they are, what they're doing, what the camera does, what the light is, what the mood and style are. With photo to video, the first items are already covered by the photo, so what's left is motion, camera and light. If the clip runs longer than a few seconds, break it down by seconds: what happens at second one, what happens at second three.
A separate note on the space around the frame. In one of my lessons I burned a lot of money on generations until I explained to the model where exactly I was sitting and what was outside the frame. As soon as I described it, the model did exactly what I needed. The AI can't see past the edges of your photo, so it makes things up. One sentence about the surroundings saves dozens of attempts.
One more thing: a short prompt preserves more detail than a long one. With a long description, the model once changed my character's appearance. I collected full formulas with examples in a separate guide: AI video prompts that the model gets right the first time.
There isn't a universal one, and I won't pretend there is. Here's how I split the work.
For storyboard frames I use two image models, Nano Banana and Seedream: often one handles a shot better than the other, often they complement each other. The key frame usually comes from Seedream, the fixes from Nano Banana.
Prices differ between models and change often, so I'll only say what I know for sure. Kling's cheapest option is a clip under five seconds. Seedance 2.5 is the most expensive model in my kit. Whatever platform you use, check the price before you hit generate.
On attempts, I'll be straight with you: the first generation is almost never the final one. In my intro lesson, the first attempt had weak motion, so I published the second. In the ad video lesson I spotted a couple of glitches in the first generation and said what I always say: by the second or third generation that problem goes away. Budget two or three attempts per shot.
Even so, the speed is on another level. A finished ad from a reworked prompt took me four minutes. Putting together a photo set for a construction timelapse takes about twenty. And clients pay for that timelapse on a completely different scale: renovation companies and event agencies buy it to show a stage being built from a single photo.
The video looks slowed down. Generators often output slightly slow-motion footage. That's why I speed up almost every shot: it looks more real, and the speed-up hides some of the distortion.
Too many references. When you feed in too many photos of your character, the result is awful. I've never used more than five.
Too few source frames for a timelapse. I mentioned the two-photo renovation already: the model fills the gaps with junk. Shoot more in-between stages.
Someone else's face. In my experience, the most realistic models, Sora and Veo, refuse photos of other real people. Your own photos are easier, and you shouldn't use other people's photos without permission anyway.
Treating the first attempt as final. The most expensive beginner mistake: the flaw ends up in front of your viewer.
The most common job is ads and product videos. Whatever you sell, use your real product: shoes, your own sneakers; jewelry, your own ring. Nobody needs a video of a product that doesn't exist.
Second is visualizing something that hasn't happened yet. My "AI Wedding" is exactly that: the wedding scenes were built with AI, and next to them in the video is the bride's real reaction. Same category: the stage-build timelapse from a single photo that event agencies order.
Third is memories: old travel photos become a clip where the AI analyzes the shots, writes the script and generates the video. No prompts needed.
Fourth is big projects. On a video for a well-known pop artist, where I was the AI art director, the lead character was fully generated with AI, and the interiors were made in a way a live-action budget would never have allowed.
For a first result, what's above is enough: one photo, one movement, an honest description of the space around the frame, and two or three attempts. If you want to go deeper into product videos, storyboards and how I build whole sequences, I post deeper breakdowns and lessons in English on Patreon.
It depends on the job. I like Kling for realism, Veo draws timelapses better, and Grok Imagine animates photos cheaper than Seedance. I use Seedance 2.5 when I need to tell the model exactly what to take from the photo and what to ignore.
Yes, and it's the fastest route: one photo plus a script written by a text model. You get less control, but for social media and testing an idea it's enough. For a clip that comes out exactly as planned, build a storyboard from several frames.
The more in-between stages, the better. With two photos of a renovation, I got generative garbage.
Many generators do that. Speed the shot up slightly in the edit: it looks more real and hides some distortion.
No. Their appearance is already in the photo. The prompt is for motion, camera and light. A long description can change your character; a short one keeps the details.
AI video generation · 12 November 2026
How to make AI videos step by step: idea, script, reference frames, generation, sound and edit. A filmmaker's workflow, plus the mistakes beginners make.
AI video generation · 12 November 2026
How to animate a photo with AI: pick the right shot, choose a video model, write a short prompt, keep the face stable and catch the flaws before you post.
Models and tools · 12 November 2026
Seedance 2.5 by ByteDance: what it can do, how it differs from 2.0, where to use it and how to prompt it. A filmmaker's take, based on my own generations.