A generated product shot comes back with your brand name nearly right: a letter shifted, the kerning off, a word that is almost the word. It looks fine at a glance and wrong to anyone who knows the brand, which is worse than no label at all.
The cause is structural and the fix is not a better prompt.
TL;DR
- Models learn the visual texture of writing, not that a glyph sequence is the only correct one
- No prompt fixes it. Naming the text in quotes helps marginally and does not solve it
- The fix: photograph the real product, cut it out at 12 credits, composite into a generated scene
- For motion: Seedance 2.5 at 157 credits holds fine detail through movement best
- Never generate the product you sell. It misrepresents what ships
Why it happens
A diffusion model has seen millions of images containing text and has learned the statistics of typography: stroke weight, spacing, how words sit on a package. What it has not learned is that a specific sequence of glyphs is the only correct output.
So it generates something with the texture of writing. On a decorative sign that is tolerable. On your product it is your brand name, approximately.
In video it is worse, because the label has to be identically correct across every frame, and drift between frames draws the eye more than a static error would.
The fix that works
Photograph the product. Composite it.
- Photograph it properly. Window light, mid-morning, plain surface, phone slightly above and in front. Twenty frames, pick one.
- Note the light direction and how hard it is. This decides whether the composite reads as real.
- Cut it out. Background removal, 12 credits.
- Generate the environment, prompting for the light you noted. Seedream 5 at 6 credits, so twenty directions for 120.
- Composite in your editor, matching light and colour temperature.
- Animate from the composite if you need motion, which keeps the photographed label in every frame.
Canon 8frame prices, a credit is $0.01 at pack rate:
| Step | Route | Credits |
|---|---|---|
| Cut out the product | background removal | 12 |
| Twenty environment directions | Seedream 5 at 6 | 120 |
| Unify the grade | Flux Edit | 7 |
| Animate, draft at 480p | Wan 3.0 | 24 |
| Animate, final holding the label | Seedance 2.5 at 720p | 157 |
| Upscale the still | Clarity | 13 |
A full product set: about 330 credits, roughly $3.30, with a label that is exactly your label.
Why Seedance 2.5 for the motion
It costs 83 percent more than Kling v3 Standard, and the narrow thing it is good at is exactly this: holding fine detail and product identity through movement. For a product clip, that is the entire job.
Even so, check every frame. Freeze on the label at several points, not just the first. Drift accumulates.
What does not work
Prompting harder. Naming the text in quotation marks and specifying position helps a little. It does not make the model reliable.
Nano Banana Pro at 21 credits is genuinely the best model here for text inside an image, and it is still not dependable enough for a brand mark. Use it for a poster headline, not for your packaging.
Editing the text afterwards. Kontext Max at 13 credits is the best of the edit tools at changing text and remains unreliable.
Upscaling. There is no video upscaler on this canvas, and on stills an upscale infers detail rather than recovering it, so a garbled label becomes a sharper garbled label.
The wider rule
Do not generate anything a customer will encounter in reality: your product, your food, your premises. A generated version misrepresents what ships, which is a returns problem, a reviews problem and in most markets an advertising standards question rather than a matter of taste.
Generate the world. Photograph the thing.
FAQ
Why is the text on my generated product wrong? The model learned what writing looks like, not what it says. No prompt fixes it.
Which model handles product labels best? Seedance 2.5 at 157 credits for motion, and photographing the product for accuracy.
Can I fix a garbled label in an edit? Unreliably. Composite the real photograph instead.
Can I generate my product at all? Not for anything a customer sees before buying.
Photograph the label, generate the room. The 8frame canvas is free and unlimited, and generation is paid from $19/month.