A packaging mockup made with AI starts from your flat art, not from a prompt: export each panel of the dieline as an image, attach the front panel (and the sides) as references to an image node, describe the box, the surface and the light, and then compare the result with the artwork panel by panel. On 8frame the chain is upload, image model with references, a written fix if needed, background removal and an upscale. The limit people do not expect: the canvas takes raster images only, so a PDF or SVG dieline has to be exported to PNG first, and no image model returns vector files.
This page is about the step between "the design is approved" and "the box exists". If you need a tool list, see the mockup generators we checked. If the box already exists and you want shots of it, start from the product photo generators page instead: a photo of the real box beats any render.
TL;DR
| Step | Node or model on 8frame | Fixed price in credits | Input it needs |
|---|---|---|---|
| Flat panels in | Upload | none | PNG, JPG, WebP, GIF, AVIF or BMP, up to 50 MB each |
| First render | Nano Banana 2 (the default) | 8 / 11 / 17 / 22 at 0.5K / 1K / 2K / 4K per image | prompt plus up to 10 references |
| Render with many panels | Seedream 5 Lite | 6 per image | prompt plus up to 14 references |
| Hero render | Nano Banana Pro | 21 at 1K or 2K, 42 at 4K | prompt plus references |
| Written fix | Flux Edit (FLUX.2 pro) | 7 for a roughly 1 MP input, more for larger inputs | the render plus one instruction |
| Second edit model | Kontext Max | 13 | the render plus one instruction |
| Cut-out | Background removal | 12 | one image |
| Enlarge | BRIA Increase Resolution | 6, 2x or 4x, up to 8192 x 8192 | one image |
Prices are per image as the app shows them; there is no fixed price for "a mockup", because it depends on how many takes you keep.
The board, node by node
There is no published template for this yet, so here are the steps in words. Lay them out left to right on one board so a colleague can follow the chain.
1. Export the panels, not the whole dieline
A dieline is a technical drawing: cut lines, fold lines, bleed, glue flaps. Image models do not read it as a box. They read it as a flat picture with odd lines on it. So export each visible face as its own image: front, one side, top. Trim the bleed. Turn off the cut and fold layers. Keep the colours as they are in the approved file.
The canvas accepts PNG, JPG and JPEG, WebP, GIF, AVIF and BMP up to 50 MB. A PDF or SVG straight from your packaging software will not upload, so export to PNG. If the art is small, enlarge it before it goes in: BRIA (6 credits) enlarges 2x or 4x and keeps the content, while Clarity (13) always enlarges 2x on 8frame and is diffusion-based, so it can add detail you did not draw. On packaging, you do not want new detail on your logo.
2. Write the box as a spec
Before you prompt, write a short spec on a sticky note next to the nodes: the shape (tuck box, sleeve, pouch, tin), the proportions (say "the box is twice as tall as it is wide"), the material (matte board, gloss, kraft, foil), the surface it stands on, the light and the angle. A spec makes the second and tenth take comparable. It also keeps the person who approves the shot honest: they check against the spec, not against taste.
3. Generate the first render with the panels attached
Wire the front panel into an image node as a reference. Nano Banana 2, the node's default, takes up to 10 references and costs 8 to 22 credits per image depending on size. Seedream 5 Lite takes up to 14, the most on the roster, at 6 credits, which helps when you attach every face plus a material swatch. Nano Banana Pro (21 at 1K or 2K) is the option for the one hero render.
Prompt the scene, not the design: "a matte white tuck box standing on a pale oak table, three-quarter view showing the front and the right side, soft window light from the left, the front artwork exactly as in the reference." There is no negative-prompt field on any node, so say what you want, not what you do not want.
We have not measured which model keeps flat artwork most faithfully. Run the same spec on two models and keep the take that matches the panels.
4. Fix one thing at a time
When the render is close, fix the rest with a written instruction instead of a new take. Flux Edit costs 7 credits for a roughly 1 MP input and rises with input size (worst case 87), so edit before you enlarge. Kontext Max (13) is the second edit model. One instruction per run: "straighten the logo so it is level with the top edge", then "remove the reflection on the right side". Several changes in one prompt make it hard to see which one broke the type.
5. Cut it out and enlarge last
Background removal (12) gives you the box alone for a white-background listing or a layout. Then enlarge: BRIA (6) for 2x or 4x up to 8192 x 8192, or Topaz (11 per started 24 MP block) for a 4x precision pass. Upscale last, because every earlier step works on the pixels you give it.
6. Put the approved art next to the render
Keep a frame on the board with the approved panel on the left and the render on the right. Every take gets this side-by-side check before it leaves the board.
Which model for which step
| Need | Model | Why this one | Limit |
|---|---|---|---|
| Many panels and a material swatch as refs | Seedream 5 Lite | 14 references, 6 credits | flat price at 2K and 3K |
| Fast first takes | Nano Banana 2 | the default, 8 at 0.5K | up to 10 references |
| One hero render for the page | Nano Banana Pro | 21 at 1K or 2K, 42 at 4K | higher price per take |
| Many reference images with a fixed quality setting | GPT Image 2 | 1 to 10 references; high (default) 29 square, 20 other ratios | price changes with ratio |
| Small fixes on a good take | Flux Edit, Kontext Max | 7 (1 MP) and 13 | Flux Edit rises with input size |
| Box on its own | Background removal | 12 | image only |
Every image model we measured returns PNG. None returns a vector file, so a render never replaces the print file. The print file stays the source of truth for the printer; the render is a picture of it.
Where it breaks
- Small type. Ingredient lists, legal lines, barcodes and recycling marks are the first things to change. A model redraws them; it does not copy them. Our page on why AI changes a product label covers the fixes. On packaging, the practical rule is: never publish a render whose small print a buyer could read and rely on. Crop it out, keep it soft, or composite the real panel back in your own editor.
- Proportions. A model may make the box squatter or deeper than the dieline. Measure: the height-to-width ratio in the render should match the spec.
- Wrap-around art. A pattern that runs across a fold can break at the edge. Attach the side panel too, and check the join.
- Finishes. Foil, spot gloss and embossing are hard to describe and easy to overstate. If the real box will be matte, do not let a glossy render sell it.
- Colour. Generated colour can shift from the approved file. Check the brand colours on the render against the panel at full size (Original in the download menu), not on a thumbnail.
When a render is not allowed as a product photo
A packaging render is a fine concept image for a pitch, a retailer deck or a pre-launch page. Some platforms limit renders where they could be taken for the real thing. Kickstarter's help centre says "hardware and product design creators are required to show working prototypes", and: "We do not allow photorealistic renderings." So if a render is going onto a crowdfunding page, read the platform's current rules first, and label the image as a render on any page where a buyer could mistake it for a photo of the finished pack.
Marketplaces have their own image rules for the main image. The safe pattern is the same everywhere: once the box is printed, photograph it and use the AI step for scenes and backgrounds around the real photo.
From photo to motion
Once you have an approved render, it can be the start frame of a short product clip. Kling v3 Standard is image-to-video only on 8frame (86 credits for 5 s with sound), so it needs exactly this kind of still. For a camera move from one image, the Video Effect tool turns one image into one short clip from one of 121 presets for 78 credits. The product video template guide covers the motion side; the same small-type warning applies to every frame.
FAQ
Can I upload my dieline PDF straight to the canvas? No. The canvas accepts PNG, JPG and JPEG, WebP, GIF, AVIF and BMP up to 50 MB. Export each visible face as a PNG with the cut and fold layers turned off, then upload the panels.
Can AI give me a print-ready packaging file? No. Every image model we measured returns PNG, and none outputs vector or SVG. Keep your packaging file as the print source; the render is only a picture of it.
Which model should I use for a box with many panels? Seedream 5 Lite takes up to 14 reference images at 6 credits per image, the most references on the roster. Nano Banana 2 takes up to 10. Run both on the same spec and keep the take that matches your artwork.
Why does the small print on my render look wrong? Models redraw text instead of copying it, so small type, barcodes and legal lines drift first. Crop them, or composite the real panel back in your own editor before you publish.
Can I use a render on a crowdfunding page? Check the platform's current rules first. Kickstarter's help centre says it does not allow photorealistic renderings for hardware and product design projects; label renders anywhere a buyer could mistake them for a photo of the finished product.
Sources
- Kickstarter Help Center, article 16236428 (dated 2026-08-06: working prototypes required for hardware and product design; no photorealistic renderings), checked 2026-11-09
- 8frame prices, defaults and limits: the 8frame app, checked 2026-11-09
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.