Correction, 2026-10-02: the first version of this page said Seedance 2.5 "holds fine detail through movement best" and was the model to use for a label in motion, called Nano Banana Pro "genuinely the best model" for text in an image and Kontext Max "the best of the edit tools" at changing text, and said that quoting the text in the prompt "helps marginally". We have not measured any of that, and no model keeps a label exact on its own, so those lines are gone and the page now tells you to check every frame. It also said "never generate the product you sell", which contradicted our own advice to restage your real product photo with an image model; it now says to start from the real photo and check the result against it. The price table used Clarity to upscale the still without saying that Clarity is diffusion-based and can redraw small text; it now uses BRIA Increase Resolution (6 credits) and says to check the label after any upscale. The Seedream step did not say that the Seedream node opens on 5 Pro. The page had no sources: the explanation of why models misspell and the line about advertising law now cite a Google Research paper and the FTC, checked today.
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.
No model keeps a label exact on its own, and no prompt guarantees it. What works is starting from the real product, checking every frame, and putting the real label back wherever the words must be exact.
TL;DR
- Image and video models don't spell reliably. A Google Research paper traces part of the problem to models that cannot see the individual letters of the words in a prompt
- No prompt guarantees the label. Write the exact text in quotation marks anyway, then read every output letter by letter
- Start from the real product photo: cut it out (background removal, 12 credits) and use it in the scene or as the reference, rather than letting a model invent the product
- Fix what drifts: regenerate, correct it with an image edit tool (Flux Edit, about 7 credits at 1 MP, or Kontext Max, 13), or composite the real label in 8frame's Video Editor or your own editor
- In motion, check every frame, not just the first: no video model keeps a label exact
Why it happens
Image models learn from huge numbers of pictures that contain text, so they learn what writing looks like: stroke weight, spacing, how a word sits on a package. Producing one exact sequence of letters, and only that one, is a different problem.
A Google Research paper by Rosanne Liu and colleagues, "Character-Aware Models Improve Visual Text Rendering", opens with: "Current image generation models struggle to reliably produce well-formed visual text." It names one contributing factor: popular text-to-image models "lack character-level input features, making it much harder to predict a word's visual makeup as a series of glyphs." In the authors' tests, models that could see individual characters rendered text better than models that could not. The paper dates from 2022 (revised 2023). Whichever model you use today, treat label text as something to check, not something to trust.
On a decorative sign in the background, a near-miss is tolerable. On your product it is your brand name, approximately.
In video it is harder still. A video model generates every frame, label included, and the label has to be right in all of them. Drift from one frame to the next draws the eye more than a static error would.
Start from the real product
Don't ask a model to invent your product. Start from a photo of the real thing.
- Photograph it. Window light, a plain surface, the phone slightly above and in front, the label facing the camera. Take several frames and keep the sharpest.
- Note the light direction and how hard it is. You will ask for the same light in the scene.
- Cut it out. Background removal, 12 credits.
- Build the scene. There are two routes:
- Composite: generate an empty environment with the light you noted, for example on Seedream 5 Lite at 6 credits (the Seedream node opens on 5 Pro, so switch to 5 Lite), then place your cutout into it in your own image editor, matching light and colour temperature. The label stays your photographed label.
- Restage: give the cutout to an image model as a reference image, for example Nano Banana Pro at 21 credits for 1K or 2K, with up to 10 reference images (the Nano Banana node opens on Nano Banana 2, so switch to Pro). The model redraws the product, label included, so read the label letter by letter. Prompts for this are in Nano Banana Pro prompts for product photography.
- Unify the grade if the product and the scene don't match. Flux Edit is about 7 credits for a 1 MP image and more for larger ones. An edit model returns a new image, so read the label again afterwards.
- Animate from the finished still if you need motion: send it to a video model as the start image.
Animating without losing the label
Send the finished still to a video model as the start image. On 8frame:
- Kling v3 Standard needs a start image and follows its shape, at 720p: 57 credits for 5 silent seconds, 86 with sound. The Kling node opens with sound on.
- Seedance goes up to 720p. The node opens on Seedance 2.5, at 157 credits for 5 seconds (70 at 480p); Seedance 2.0 is 108 at 720p and 48 at 480p.
- Veo 3.1 Fast is the 1080p option. From a start image it always makes 8 seconds, in 16:9 or 9:16, for 168 credits with sound.
We have not measured which of these keeps a given label steadiest, and none of them keeps it exact. To choose, run the same still through two models at a cheap setting, for example Seedance 2.0 at 480p (48 credits) and Kling v3 Standard without sound (57), then freeze on the label at several points in each clip: the start, the middle and the end. Keep the label facing the camera, keep rotation small and keep the move slow.
When a frame still gets the label wrong:
- Regenerate. A finished clip you reject still costs its credits; only a run that fails is refunded automatically (do failed AI generations cost credits?).
- Composite the real label over the clip. In 8frame's Video Editor you can upload your own image and place it on the timeline as an image layer over the clip. A static layer suits a shot where the product stays put. For a product that moves, composite in your own editor, or turn the label away from the camera in that shot.
What to try on a still, and what to check
Name the text in the prompt. Put the exact words in quotation marks and say where they sit. It does not guarantee the spelling, so read every output letter by letter.
Use your photo as a reference. On Nano Banana Pro (21 credits at 1K or 2K) the real product photo can go in as a reference image. The model still redraws the label, and we have not measured how often any model spells it right.
Edit the text. You can ask an image edit tool to correct a word: Flux Edit (about 7 credits at 1 MP) or Kontext Max (13). Check the corrected word and the rest of the label, because the edit can change more than the word you asked about.
Regenerate. Run it again and compare. A finished image you reject still costs its credits.
Composite the real label. When the words must be exact, place the real label from your product photo over the generated one in your editor. It is the only route here where the label on screen is your own photograph.
Don't expect an upscale to fix spelling. There is no video upscaler on 8frame. On a still, an upscaler enlarges what is there and cannot know your brand name. Clarity (13 credits) is diffusion-based, so it can redraw small text; BRIA Increase Resolution (6 credits, 2x or 4x) is the other option on the upscale tool. Read the label after any upscale.
The price of a product set
Prices on 8frame; a credit is about $0.01 at pack rate.
| Step | Route | Credits |
|---|---|---|
| Cut out the product | Background removal | 12 |
| Twenty environment directions | Seedream 5 Lite, 6 each | 120 |
| Unify the grade | Flux Edit, 1 MP image | about 7 |
| Two test clips from the same still | Seedance 2.0 at 480p (48) and Kling v3 Standard, silent, 720p (57) | 105 |
| Final clip, 5 s at 720p | Seedance 2.0 (108; the node opens on 2.5 at 157), or keep the Kling clip | 0 to 108 |
| Enlarge the still | BRIA Increase Resolution, 2x or 4x | 6 |
A full product set comes to about 250 to 360 credits, roughly $2.50 to $3.60 at pack rate, before reruns. The label on the still is your photographed label; in the clips, check it frame by frame.
The wider rule
Don't let a model invent anything a customer will meet in reality: your product, your food, your premises. If the ad shows something the customer doesn't receive, that is a returns problem, a reviews problem and a legal question: in the US, the FTC says that "claims in advertisements must be truthful, cannot be deceptive or unfair, and must be evidence-based."
Generate the world. Photograph the thing.
FAQ
Why is the text on my generated product wrong? The model generates the label as part of the picture, and image and video models don't spell reliably. A Google Research paper links part of this to models that cannot see the individual letters of the words in a prompt. No prompt guarantees the spelling.
Which model keeps a product label exact? None, on its own. Start from the real product photo, check every frame, and composite the real label where the words must be exact. To choose a video model, run the same still through two of them at a cheap setting and compare the label frame by frame.
Can I fix a garbled label in an edit? Sometimes. Try Flux Edit (about 7 credits at 1 MP) or Kontext Max (13) and read the result letter by letter. When the words must be exact, composite the real label from your photo instead.
Will upscaling fix a garbled label? No. An upscaler enlarges what is there, and a diffusion-based one such as Clarity can redraw small text. There is no video upscaler on 8frame.
Can I generate my product at all? Restage it from your real product photo rather than from a text description, and check the result against the real product before anyone sees it: the label, the shape and the colour.
Sources
- Rosanne Liu and colleagues, Google Research, "Character-Aware Models Improve Visual Text Rendering" ("Current image generation models struggle to reliably produce well-formed visual text"; popular text-to-image models "lack character-level input features, making it much harder to predict a word's visual makeup as a series of glyphs"; character-aware models outperformed character-blind ones on text rendering; first posted December 2022, revised May 2023): arxiv.org/abs/2212.10562, checked 2026-10-02
- US Federal Trade Commission, Advertising and Marketing ("Under the law, claims in advertisements must be truthful, cannot be deceptive or unfair, and must be evidence-based"): ftc.gov/business-guidance/advertising-marketing, checked 2026-10-02
- 8frame prices, resolutions and settings (background removal, Seedream 5 Lite and 5 Pro as the Seedream node's starting model, Nano Banana Pro with up to 10 reference images and Nano Banana 2 as the node's starting model, Flux Edit priced by image size and Kontext Max, Kling v3 Standard and its start-image requirement, Seedance 2.0 and 2.5, Veo 3.1 Fast and its 8-second lock with a start image, Clarity and BRIA Increase Resolution, no video upscaler, automatic refunds for failed runs): 8frame tool registry and cost functions, read 2026-10-02
- Video Editor image layers from your own uploads: 8frame front-end code, read 2026-10-02
Photograph the label, generate the room.
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.