To make many static ads from one product, lay the board out as a matrix: rows are angles (the reason to buy), columns are audiences or formats, and every cell is one image node with your product photo attached as a reference and a prompt that changes only what that cell is about. The model you pick decides how well the product survives, because the models differ in how many reference images they accept: from none on FLUX.2 [pro] to 14 on Seedream 5 Lite. The limit people do not expect: 8frame has no batch node and no variable injection, so a matrix of 12 cells is 12 nodes you duplicate and edit by hand, and one account runs at most 10 generations at a time on Starter (20 on Creator, 50 on Agency).
This page covers both kinds of variation: new concepts (different angle, scene and message) and variations of one concept (same idea, new scene or crop). They do different jobs, and a good board holds both.
TL;DR
| Step | Node or model on 8frame | Price in credits | Input |
|---|---|---|---|
| Product reference | upload, plus Background removal if busy | 12 per image for the cut-out | Your real product photo |
| Concept stills, default | Nano Banana 2 | 8 / 11 / 17 / 22 per image (0.5K / 1K / 2K / 4K), up to 4 per run | Up to 10 references |
| Concept stills, many references | Seedream 5 Lite | 6 per image | Up to 14 references |
| Stills with text in the picture | GPT Image 2 | high: 29 square, 20 other ratios | 1 to 10 references |
| Scene edit of an approved still | Flux Edit (FLUX.2 pro) | 7 for about a 1 MP input | The approved still |
| 4:5 feed version | Nano Banana or Kontext Max | Kontext Max 13 | The approved still |
| Bigger file for print or zoom | BRIA Increase Resolution | 6, 2x or 4x | The approved still |
Concepts versus variations
A variation keeps the idea and changes the surface: the same "morning routine" shot in three bathrooms, or the same layout with a new background colour. A concept changes the idea: "morning routine" versus "gift for her" versus "travel size fits the bag". Variations tell you which execution of an idea works; concepts tell you which idea works. If every cell on your board is a variation of one idea, you learn little about what to say.
So the board below is a concept matrix first. Variations come later, only for the concepts that earn them.
The board, node by node
The concept-matrix template is planned for 8frame's templates, not live yet. Build it on an empty board from the 8frame canvas, or adapt a workflow template.
1. The product reference
Upload the best photo of the real product, front-on, even light. If the background is busy, run Background removal (12 credits) and keep the cut-out next to the original. Add a second angle if the product has a back or a side that matters. These are the references every cell will use. The product is the one thing that must not change between cells.
2. Draw the matrix with frames and sticky notes
Use frames for rows and sticky notes for the labels. A simple matrix: four angles down the side (problem it solves, who it is for, how it is used, why it is different) and three columns across (a person using it, the product alone in a scene, a flat graphic layout with room for text). Twelve cells. Write each cell's one-line brief on a sticky note in the cell.
3. One image node per cell
Make the first cell's node: product references attached, prompt built from the cell's brief, ratio set to the placement (9:16 for Reels and Stories, 4:5 for feed; see the next section). Get that one right. Then duplicate it into the other cells and change only the prompt. 8frame has no batch node and no variable fields that fill prompts from a table, so this is the manual step; keeping one tested node as the source of the others is what keeps the settings identical.
Nano Banana 2 and GPT Image 2 return up to 4 images per run, so each cell can give you four takes of its concept at once. One account runs at most 10 generations at a time on Starter, 20 on Creator and 50 on Agency. On Starter a 12-cell board fills the queue: start the next cells as the first ones finish.
4. Review the matrix, not the images
Look at the board as a grid. Kill whole rows, not single images: if the "why it is different" row is weak in all three formats, the angle is weak. Mark winners with a colour on the frame.
5. Variations of the winners
For each surviving concept, make variations on the approved still: a new scene with Flux Edit (7 credits for about a 1 MP input, rising with input size), a 4:5 version on Kontext Max (13) or Nano Banana, a larger file with BRIA (6, 2x or 4x). Now the variations test execution of an idea you already know works.
Which model for which step
The nuance is references: how many product pictures a model can hold on to, and what each one costs.
| Model | Reference images | Ratios include 4:5? | Price on 8frame | Use it for |
|---|---|---|---|---|
| Seedream 5 Lite | up to 14, the most on the roster | No | 6 | Cells that need product, scene and style references together |
| Nano Banana 2 | up to 10 | Yes | 8 / 11 / 17 / 22 by resolution | The default for most cells; 4:5 feed images |
| Nano Banana Pro | up to 10 | Yes | 21 at 1K and 2K, 42 at 4K | Larger files where you want the Pro model |
| GPT Image 2 | 1 to 10 | No (1:1, 16:9, 9:16, 4:3, 3:4) | high 29 square, 20 other ratios | Layouts with text in the image |
| Qwen Image 3 | 1 to 3 for edits | No | 6 at 1K, 11 at 2K, 1 to 6 per request | Cheap drafts of many cells |
| Grok Image 2.0 | up to 3 | No | 9 at 1K (6 low), up to 15 at 2K with 3 references | Quick alternates |
| FLUX.2 [pro] | none on this node | No | 5 at 1 MP | Backgrounds and textures without the product |
Two details from that table change how you build. GPT Image 2 is the only image model on 8frame that prices by aspect ratio: at its default high quality, a square image is 29 and any other ratio is 20, so for a 9:16 matrix it costs less per image than square. And FLUX.2 [pro] takes no reference image on its generation node, so it cannot be the product cell; use it for plain scenes you composite into later, or use Flux Edit on an existing image instead.
None of these numbers say which model keeps a product most faithfully; we have not measured that. Compare every take with the product photo: shape, colour, label text, logo.
Placement sizes
Feed images on Meta are 4:5 and Stories are 9:16; the sizes, file limits and ratio tolerances per placement are on Meta ad specs by placement. What matters for this board: of the image models above, only Nano Banana (and the Kontext Max edit) offer a 4:5 setting, so set the 4:5 cells there. If a download is smaller than the size you need, upscale the approved still with BRIA (6 credits) instead of generating every draft at the most expensive resolution.
Text in the picture
Keep headline text out of most cells and add it in the ad tool or in your design tool, where it is editable and exact. Image models redraw letters, and a typo in a price is the kind of mistake that runs for a week. When a cell needs text in the image (a flat graphic layout), use GPT Image 2 and read every word on the result.
Where it breaks
- No batch node. Each cell is a node you duplicate and edit. There is no variable injection, no output formatter, no export preset.
- A limit per plan. One account has at most 10 active generations on Starter, 20 on Creator and 50 on Agency.
- The product drifts. Labels, logos and colours can change between takes. See why AI changes a product label.
- Not every model takes references. FLUX.2 [pro] generation takes none; Qwen Image 3 and Grok Image take up to 3.
- 4:5 is rare. Only Nano Banana and Kontext Max have it among the image models.
- No vector output. No image model on 8frame outputs SVG. Logos and type for print come from your design files.
FAQ
How many ad variations should I make from one product?
Start with concepts, not counts: four angles in three formats is a twelve-cell matrix. Make variations only of the concepts that survive review.
Which AI model keeps my product the same across variations?
Choose by reference capacity and check each take: Seedream 5 Lite accepts up to 14 references, Nano Banana 2 and GPT Image 2 up to 10. We have not measured which keeps a product most faithfully.
Can I generate all the variations in one click on 8frame?
No. There is no batch node. Duplicate one tested node into each cell and change its prompt; Nano Banana 2 and GPT Image 2 return up to 4 images per run.
What size should a static Meta ad be?
Meta recommends 4:5 at 1440 x 1800 for feed images and 9:16 at 1440 x 2560 for Stories.
Should the headline be in the AI image?
Usually not. Add text in the ad tool or a design tool. If a layout needs text in the picture, use GPT Image 2 and proofread it.
Sources
- Meta Ads Guide, Instagram Feed image (4:5, 1440 x 1800, 30 MB, 1% tolerance), checked 2026-11-09
- Meta Ads Guide, Facebook Feed image (4:5, 1440 x 1800, 3% tolerance), checked 2026-11-09
- Meta Ads Guide, Instagram Stories image (9:16, 1440 x 2560), checked 2026-11-09
- 8frame prices, reference limits, ratio settings, 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.