Guide · 8 min read · Updated November 3, 2026
Get great frames, and fix them in words
The photos to give, how to check a frame, how to show a product that folds or opens, and which model makes which frame.
Why frames come first
Every Genien video starts as still frames, one or more per shot. A frame costs about a fifth of a clip, and the clip follows its frame closely, so the cheapest place to get a video right is its frames. This is how professional designers work with AI images, said for a seller without a designer.
Set up the product, with the right photos
Start on the Products tab. Paste the product's page, and Genien reads it and every picture on it, then asks up to five questions the page doesn't answer, such as which packaging is sold today and how the product is really used. Add the photos and a phone clip, read the short brief, confirm the claims Genien found and approve it. Every idea, frame, clip and check follows that brief, so a drink is never shown eaten with a spoon, and an ad says only the claims you allowed, with their fine print, while label lines like "Low GI" stay unreadable.
- 3 or 4 photos of the product alone on a plain background: the front, a side and the back.
- One photo of it in a hand, so the models get its size right.
- If it folds, opens or changes shape, one photo of each state, such as folded and open. Genien uses them for the first and last frames of that shot.
- Sharp photos taken from a little distance. A phone held close bends straight edges, and the models copy the bend.
- No collages, infographics, size charts or room scenes. Several different pictures read as several different products.
- To show the product working, film it on your phone for 5 to 10 seconds in one steady take. No AI model knows how your product works, and a real clip is the strongest proof.
Make the storyboard first
Choose Storyboard first when you run a video, or make it your default in Settings. Genien makes only the frames, in about two minutes and for a fraction of the price. Check them, fix any you don't like, then press Make the video. The clips start from the frames you approved.
Check each frame in this order
- The product.
- The same shape, parts and colors as your photos. Read the label letter by letter and look at the logo. Exactly one product, at the right size for the scene.
- The physics.
- It rests on something, with a shadow that matches the light. Nothing floats or looks pasted on, and straight lines stay straight.
- The story.
- The frame shows what the voice says at that moment.
- The phone screen.
- Captions cross at about 60% of the height and the app's buttons cover the bottom third, so the subject sits above them.
- The AI look.
- Skin like plastic, colors too vivid, everything too clean. Real photos have texture, small flaws and one clear light.
Genien checks every frame against your photos before you see it, but a check misses things a person sees at once. Your look is the last word.
Fix a frame
- A detail is wrong and the rest is good: tell the chat what to change, with the frame's @name. Genien edits that picture and keeps everything else, the way designers fix a frame. One change per message.
- The whole frame is wrong: ask the chat to rewrite it, or describe the frame again in your own words.
- Each frame is made twice and the better one goes on. The other one is in the frame's Versions, a click away.
Say exactly what you see:
| Vague | Clear |
|---|---|
| Make it better | Make the grout lines straight, parallel and the same width |
| Fix the hand | The right hand holds the bottle with five fingers, thumb in front |
| Nicer light | Warm window light from the left, with a soft shadow to the right |
| The product looks off | The lid is gold, not silver, and the label faces the camera |
| Less fake | Natural color, visible skin texture, no glossy retouching |
Show a change, such as folding or opening
AI video models don't know how your product works. Given one frame and asked to unfold a cabinet, they guess, and the parts bend or melt. So Genien plans a change as a before and after:
- The first frame shows the product before, such as folded flat.
- The last frame is made from the first one as an edit: the same camera, light and room, with only the product changed.
- A change with several steps also gets a middle frame, such as half unfolded.
- The clip moves between the frames, with the camera locked and each step named in order.
You check every frame of it in the storyboard before the clip is made, and photos of each state help most. A simple change, such as a lid opening, usually works. A mechanism with many parts can still come out wrong, and then a cut between the two frames or your own clip does the job.
Write a frame yourself
- Write full sentences in this order: what the frame is, the subject and what it does, the place, the camera, then the light.
- Mention the product as @product, so its photos go along with the words.
- Say what is there rather than what isn't: "a bare floor" works better than "no clutter".
- Name detail as something you could check: "lines of equal width", not "clean lines".
- Leave out words like 8k, masterpiece and hyper-realistic. They make a picture look generated, and Genien takes them out.
- Keep words out of the picture. Captions and the end card are set in a font, so they are always spelled right.
- For an edit, name one change, then what stays: "Change only the cabinet: it now stands open. Keep everything else the same."
A vertical 9:16 photo of @product standing on a pale oak floor next to a white wall. A low view at product height with a 50mm lens look, the whole product in sharp focus. Soft window light from the left and a soft shadow to the right.
A close vertical photo of a right hand holding @product, thumb in front and four fingers behind, over a softly blurred bathroom counter. Five natural fingers with visible knuckles. Soft daylight from a window on the left.
A candid phone photo of a woman in her thirties in a bright kitchen, holding @product near her face and smiling mid-sentence. Arm's-length front camera, a slightly bright window behind her, real skin texture, everyday things on the counter.
An extreme close-up of the tiled floor after cleaning: straight, parallel grout lines of equal width, every tile flush. Raking light from the left shows the surface, in sharp focus across the frame.
A wide vertical photo of an empty small bedroom with a pale oak floor and white walls. A bare floor in the middle foreground, soft morning light from a window on the left. Eye level, 24mm lens look, straight vertical lines.
Which model makes which frame
Genien picks the model for each frame from what it shows, and every frame names its model. You can switch the model on any frame, and the check works the same way.
| Frame | Model | Why |
|---|---|---|
| The product on its own | GPT Image 2.5 Flare | First on the 2026 leaderboards for product shots, and keeps the product closest to its photos. |
| A hand with the product | GPT Image 2.5 Flare | First for photoreal hands and people, and gets fingers and objects right. |
| A person | GPT Image 2.5 Flare | First for portraits and photoreal people. |
| A scene, or the problem | GPT Image 2.5 Flare | First for photoreal scenes. |
| A close-up where detail must be exact | GPT Image 2.5 Sunburst | The most detailed GPT Image: seams, fine print and small parts stay sharp and exact. |
| The later frames of a change, and fixes | GPT Image 2.5 Sunburst | First at edits that change one thing and keep the rest of the picture. |
| The empty place the video happens in | Nano Banana 2.1 | An empty room for the other frames to follow. A cheaper model draws it as well. |
| Every frame of a phone-style UGC video | Nano Banana 2.1 | Phone footage: Nano Banana 2.1 pictures read most like a phone photo, where GPT Image looks more polished, and one model for every frame keeps the look the same. |
From the public leaderboards of blind votes by people (LMArena and Artificial Analysis, October 2026) and our own runs. The table changes when the models do.
When something goes wrong
| What you see | Why | What to do |
|---|---|---|
| Two or three products in one frame | The photos show different items, or a collage reads as several | Keep only clean photos of the one product. Genien already asks for exactly one. |
| A garbled label or logo | Image models still misspell small text | Make it a close-up, which goes to the most detailed model, or turn the label away |
| The product changes from shot to shot | The photos disagree, or a frame was made without them | Use 3 or 4 photos of the same unit, and check that every frame mentions @product |
| A strange hand | Hands are still the most common failure | Name the hand and the grip, or frame the shot without a hand |
| It looks fake | Too smooth, too vivid, too perfect | Ask for natural color, one light source and real texture |
| The frames don't look like one video | The light or the room changed | Keep The place node wired, and fix a frame in the chat instead of starting it over |
| The fold or the opening melts in the clip | The model guesses how the parts move | Add a middle frame, cut between the two frames, or film it |
What AI can't do yet
- Show how a mechanism really works, step by step, every time.
- Copy small printed text and logos exactly. In a 2026 test of 850 products, the best model drew the whole product right in 29% of tries. That is why every frame is checked, and why your look matters.
- Know your product's size and how it's used, unless the photos show it.
Genien puts an AI label on its videos by default. Ad platforms and some US states ask for one, and it keeps your ad account safe.
Sources: the LMArena and Artificial Analysis image leaderboards (October 2026), Photoroom's test of 850 products (August 2026), the IAB's AI disclosure framework (2026), the prompting guides of Google, OpenAI, Black Forest Labs and ByteDance, and our own run log.
Next guideBrief your product so AI video gets it right