A behind-the-scenes case study documenting the real workflow, mistakes included, behind one AI-generated video made with Zoviz Canvas.
The first two attempts were bad. The guy's face changed between panels like he had a twin filling in. A dog appeared in one frame, then vanished in the next. At one point the lip-sync landed on the wrong panel, so a silent man mouthed words while the line played over someone sipping coffee.
Figuring out why took an afternoon; rewriting the prompt wasn't the fix.
Here's the workflow that finally worked: a starting image from Zoviz AI Image Generator, brought to life in Zoviz Canvas. The mess is left in because it's the useful part.
The Actual Problem
Video models are good at moving things. They're not nearly as good at inventing things while also moving them.
Asking the model to design a character, build a set, and animate a scene all in one shot means asking it to make a hundred small decisions at once. Each one is a chance for something to drift off-model.
The turning point wasn't a tutorial. It was the third attempt, where the image had already been built separately and Canvas was only asked to animate it, which worked dramatically better than the first two.
Design first, animate second. Don't ask one prompt to do both jobs at once.
Step 1: Build the Image Before You Touch Canvas
One three-panel image was generated first using Zoviz AI Image Generator: same man in all three poses, black and white editorial look, glass house with a pool behind him.
Think of it as the anchor. If his hairstyle, the lighting, and the framing of the pool weren't already locked in correctly in this still image, no amount of clever prompting later was going to fix it.
A few things worth knowing before you start:
- Generate the image at the highest resolution the tool allows; small flaws in a still get amplified once motion is added.
- Keep the lighting direction the same across every panel in the reference image itself. Don't count on the video prompt to fix mismatched shadows later.
- If a dog, a second person, or any other element is going to appear anywhere in the video, put it in the reference image from the start.
Ready to build your own anchor image?
In the character, lighting, and composition before animating anything.
Step 2: The Motion Prompt Is a Different Animal
Writing a prompt for a still image describes a moment. Writing one for video has to describe a timeline.
It needs to explain what happens first, what happens next, and what stays completely still while something else happens.
The first draft tried to get all three panels moving at the same time. That's the version where the dog started talking, for reasons that remain genuinely unclear.
Once the prompt was rewritten to treat it like a shot list for a short film, one panel active, everything else frozen, then a crossfade, things calmed down considerably.
Here's how the sequence broke down:
| Panel | Action | Talking? |
|---|---|---|
| Panel 1 | Seated, hand under chin, slow blink, eyes drift to the side | No, sips coffee |
| Panel 2 | Runs a hand through his hair, smiles at the camera | No, blinks only |
| Panel 3 | Plays with a golden retriever, dog reacts naturally | Yes, delivers the line |
| Final 2 seconds | All three panels move together as one shot | Only Panel 3 speaks |
One panel moves, the rest freeze; that single instruction did more work than anything else in the entire prompt.
Skip it and the model tends to animate a little of everything at once, which reads as noise instead of an intentional beat.
One video was just the start.
See how to build a full cinematic AI video ad campaign from scratch, strategy, prompts, and creative direction for every scene in our blog post.
Getting the Audio Right
This is the part that's easy to forget entirely.
Without explicitly specifying no vocals, and lip-sync only on the speaking panel, faint mouth movement can show up on characters who aren't supposed to be talking. It's genuinely unsettling to watch back.
A continuous instrumental track (piano, ambient pads, warm bass) that fades slightly under the dialogue, instead of cutting out, solved this.
The Negative Prompt Isn't Optional
The negative prompt is often treated as polish.
It's not polish; it's damage control for problems you can't fully predict ahead of time.
In this case, that meant excluding:
- A second person occasionally wandering into frame
- The man's face shifting slightly between panels
- A perfume bottle showing up that nobody asked for, more than once
Once the perfume bottle was explicitly excluded, it mostly stopped. It reappeared once more in a later test, for reasons that remain unclear.
Want to skip the trial and error? Bring the reference image straight into Zoviz Canvas and start from the prompt structure below instead of building it from scratch.
The Full Prompt
B&W luxury fashion editorial video, one continuous seamless scenario, consistent lighting, tone, and motion style throughout, no visual break between segments. 85mm lens, shallow DOF, soft daylight, film grain.
Three panels visible from frame one, same man, same hairstyle and face in all three, no variation. Glass house with pool, unchanged.
One panel animates at a time, others frozen.
Panel 1 (seated, tank top): hand under chin, blink, eyes drift away.
Crossfade.
Panel 2 (white T-shirt): runs his fingers through his hair casually, smiles warmly at the camera.
Crossfade.
Panel 3 (white T-shirt, no suit): plays with the golden retriever dog with his hand, dog reacts naturally, smiles.
Final 2 seconds, all panels move together, continuing smoothly as one flowing scene.
Panel 1 sips coffee, no talking. Panel 2 only blinks, lips closed, no talking. Only panel 3 speaks, turns to camera, says "Little details. Big confidence." with lip-sync.
Soft lounge instrumental, piano, ambient pads, warm bass, no vocals, fades under speech, continuous track with no cut or restart.
Minimal camera push-in, slight handheld. No subtitles, no text.
Negative Prompt:
extra people, morphing, identity change, inconsistent hairstyle, color image, text overlay, subtitles, watermark, distorted hands, two people talking, two mouths moving, lip movement on panel one or two, panels fading in later, perfume bottle, spraying motion, suit, blazer, scene split, jump cut, disconnected segments, abrupt style change
Copy it, then swap the wardrobe and dialogue line for the intended scene. The structure underneath is what does the actual work.Things That Would Change Next Time
- Render in shorter beats; anything past a few continuous seconds gave the model more room to drift. Cleaner results came from thinking in 2 to 3 second chunks.
- Stop over-describing emotion; "smiles warmly" worked better than a paragraph describing exactly how happy he should look.
- Test the negative prompt on its own first. See what the model does by default, then build the exclusion list from actual observed problems.
- Save the reference image. If the dialogue needs to change later, or the clip needs resizing for a different platform, starting from the same anchor keeps everything consistent.
Does This Only Work for Fashion Content?
No.
The 'lock the image, then animate' logic isn't specific to this kind of shoot.
It applies just as well to:
- Product shots that need to rotate without warping
- Talking-head clips where only part of the frame should move
- Multi-scene brand stories where consistency across cuts matters more than flashy camera work
The rule of thumb: the more consistency a project needs, the more time belongs on the still image before the motion prompt ever gets opened.
Ready to try a new project? Head to the our Video AI Generator to turn a reference image into a finished, cinematic clip.
FAQs
Do you have to start with a generated image, or can you animate a photo you already have?
Either works, but a generated image gives more control since panels and focal points can be designed with animation already in mind.
How much comes down to the prompt versus the source image?
More than expected. A messy reference image shows every flaw once it moves, so fixing the image first means the prompt has less to compensate for.
Is there a limit to how many things should move in one shot?
Yes. One clear focal motion at a time, with everything else pinned down, consistently beats animating a whole scene at once.
What's the most common mistake with negative prompts?
Being too generic. "No distortions" tells the model almost nothing; "no lip movement on panel one or two" tells it exactly what to avoid.
A sweet bite, zero guilt, and zero film crew.
See how FitBite Pastry built its promo video with AI in this blog post.
