Meta Andromeda is the retrieval stage of Meta's ad system: Meta's engineering post (December 2, 2024) describes it as "selecting ads from tens of millions of ad candidates into a few thousand relevant ad candidates" before ranking. Meta's business pages tie it to creative diversification, which Meta defines as "creating a wide range of ad creatives with different themes, messages, and visuals to cater to diverse audience segments", and one of them recommends "at least 20 diversified ads in Advantage+ shopping campaigns." On an 8frame board that becomes a concept matrix: one row per audience need, one column per format, each cell its own node. The limit people do not expect: 8frame has no batch node, so twenty concepts are twenty nodes, and the "4 images per request" option on image models gives you four takes of one concept, not four concepts.
What Meta says, and what it does not
Meta's own pages, quoted (all checked 2026-11-09):
- Engineering at Meta, December 2, 2024: Andromeda selects from "tens of millions of ad candidates" down to "a few thousand"; the number of ad creatives "is expected to grow significantly" with Advantage+ and generative AI; the deployed model "has achieved +6% recall improvement to the retrieval system" and "+8% ads quality improvement on selected segments."
- Meta for Business, March 27, 2025: "as businesses upload more and more creatives to support a diversification strategy, Meta Andromeda works behind the scenes to power more complex models that allows Meta to pick the right creative".
- Meta for Business, November 19, 2024 ("Three steps to optimize your performance with creative diversification"): "Launch new ads every week"; keep "at least 20 diversified ads in Advantage+ shopping campaigns"; put "20-30% of your overall ad budget towards testing new ads"; build personas with different needs and match messages to each.
- Meta for Business, April 22, 2025: creative diversification is "creating a wide range of ad creatives with different themes, messages, and visuals to cater to diverse audience segments."
- Engineering at Meta, October 28, 2026: "generative AI creative tools have helped advertisers rapidly create hundreds of millions of new ads"; the new system "surfaces the top candidates at over one billion ads per second", and "Since 2025, Meta Andromeda has delivered a 2.4% increase in conversions on Facebook."
What we did not find on any Meta page we read: a statement that Meta groups visually similar ads, a similarity score, or an "Entity ID" for near-duplicates. Those mechanisms appear in agency blogs and video summaries. One widely read practitioner blog (jonloomer.com, October 20, 2025) puts the practical point without claiming a mechanism: "Unless they're distinctly different enough, any variation is unlikely to make a meaningful impact on performance", and text differences "should be significant, not subtle." That is consistent with Meta's own wording about different themes, messages and visuals. Treat the rest as theory.
TL;DR
| Step | Node or model on 8frame | Fixed price (credits) | Input it needs |
|---|---|---|---|
| 1. Personas and messages | sticky notes on the board | none | your customer research |
| 2. Concept stills, one per cell | Nano Banana 2 | 8 / 11 / 17 / 22 per image at 0.5K / 1K / 2K / 4K | product photo as reference |
| 2b. Cheap concept drafts | Seedream 5 Lite or FLUX 3 | 6; 4 per 1K image | prompt, references |
| 2c. Text-heavy statics | GPT Image 2 | high: 29 square, 20 other ratios; medium 8 / 6 | prompt, 1 to 10 references |
| 3. Short video concepts | Kling v3 Standard | 17.01 per second with sound, 11.34 silent | one still |
| 3b. Talking concept | Pruna Avatar lip sync | 3.375 per second at 720p | portrait + audio |
| 4. Final edits | Video Editor | not a generation | the assets |
Prices are per node run. How many concepts you make is your call; Meta's page recommends at least 20 for Advantage+ shopping campaigns.
The board, node by node
There is no concept-matrix template on 8frame yet. Build it on a blank board with sticky notes (five colours) and frames.
Step 1: the matrix on sticky notes
Rows are audience needs, not demographics. Meta's own example is that younger shoppers may respond to deals while older shoppers may prefer product reviews and demonstrations. Write four or five rows, each a need and its blocker:
- "Wants it cheaper" (blocker: price)
- "Not sure it works" (blocker: proof)
- "Buying a gift" (blocker: will they like it)
- "Has a competitor product" (blocker: switching)
Columns are formats: a static image, a 5 to 10 second product video, a talking-head clip, a text-led static. Every cell is one concept: a different message in a different visual. Colour the sticky note by row so the board stays readable.
The test of a cell: if you swapped its background colour, would it become another cell? If yes, it is not a different concept.
Step 2: concept stills
One node per cell. Feed each from the same product photo so the product is consistent, and change everything else: scene, person, angle, message.
- Nano Banana 2 is the Nano Banana default: 8 credits at 0.5K, 11 at 1K, 17 at 2K, 22 at 4K, up to 4 images per request, up to 10 references.
- Seedream 5 Lite is 6 credits per image and takes up to 14 references. FLUX 3, the FLUX default on new nodes, is 4 per 1K image and takes a reference image. Use these for rough drafts of many cells.
- GPT Image 2 for text-led statics: high quality (the default) is 29 credits square and 20 for other ratios, medium is 8 square or 6 other. It is the only image model on 8frame whose price depends on the aspect ratio.
- Nano Banana Pro, 21 at 1K or 2K, for the final version of a cell that wins.
A request for 4 images gives four takes of the same prompt. Pick the best one; do not ship all four as four ads. They are exactly the near-duplicates the diversification advice warns about.
Step 3: video and talking cells
For video cells, connect the cell's still to an image-to-video node: Kling v3 Standard runs 3 to 15 seconds at 17.01 credits per second with sound, 11.34 silent; Seedance 2.0 is 108 at 720p. For a talking cell, connect a portrait and a voice track to a lip-sync node; the nodes and their prices are compared in AI lip sync price by model.
Nodes run in parallel up to the account's limit of active tasks: 10 on Starter, 20 on Creator and 50 on Agency. On Starter a 20-cell matrix runs in two waves; on Creator or Agency it runs in one.
Step 4: finish in the Video Editor
Add the message as text and captions per cell in the Video Editor. Keep one message per ad: a cell that tries to answer two needs belongs in neither row.
When a cell wins, change one thing at a time inside that row (a new hook, a new scene) and keep the other rows running. That is the weekly cadence Meta's page describes: "Launch new ads every week".
Which model for which step
| Cell type | Model | Price | Shape options | What to watch |
|---|---|---|---|---|
| Lifestyle static | Nano Banana 2 | 11 at 1K | set on the node | product scale and label |
| Draft static | Seedream 5 Lite | 6 | set on the node | rough; redo the winner on a stronger model |
| Text-led static | GPT Image 2 | 29 square, 20 other (high) | price depends on ratio | read every word |
| Product video | Kling v3 Standard | 17.01 per second with sound | follows the still | 720p |
| Product video, 1080p | Kling 2.6 Pro | 95 for 5 s with sound | own aspect setting | 5 or 10 s |
| Talking head | Pruna Avatar | 3.375 per second at 720p | keeps the portrait's shape | we quote price and settings, not results |
| Join cells into a cut | Video Editor | not a generation | timeline | the Montage node drops sound |
Where it breaks
- Colour swaps. Same scene, different background colour: one concept, not two. Change the message and the visual.
- Four takes shipped as four ads. The 4-per-request option is for choosing, not for diversity.
- Twenty nodes by hand. There is no batch node, no variable injection and no export preset node. Plan the matrix before you generate, and use "Duplicate workflow" to copy a finished board, with or without outputs, when a new product needs the same matrix.
- Text in images. GPT Image 2 is the text-led option, but every word still needs reading. Long copy belongs on the Video Editor timeline.
- Silent joins. The Montage node outputs no audio track. Join talking or music cells on the Video Editor timeline.
- Unlabelled AI. Meta labels ads made with its own AI tools and says it will also begin detecting ads made with third-party AI tools. Expect the "AI info" label.
FAQ
What is Meta Andromeda? The retrieval stage of Meta's ad system, which narrows tens of millions of candidate ads to a few thousand before ranking, according to Meta's engineering post of December 2, 2024.
How many creatives does Meta recommend? Meta's page on creative diversification (November 19, 2024) recommends "at least 20 diversified ads in Advantage+ shopping campaigns" and launching new ads weekly.
Does Andromeda punish similar ads? We found no Meta page that says so. Meta's pages ask for different themes, messages and visuals. Claims about similarity scores or grouping come from agencies and are not confirmed by Meta in what we read.
Can 8frame generate 20 variations in one click? No. There is no batch node. Each concept is its own node, and up to 10 tasks run at once on Starter, 20 on Creator and 50 on Agency.
Which image model is cheapest for drafts? Pruna p-image at 1 credit per image, text only, with no reference image. With your product photo as a reference: GPT Image 2 at low quality is 2 credits, FLUX 3 is 4 per 1K image, Seedream 5 Lite 6.
Sources
- 8frame prices, defaults and limits: the 8frame app, checked 2026-11-09.
- Engineering at Meta, "Meta Andromeda: Supercharging Advantage+ automation..." (2024-12-02): https://engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/, checked 2026-11-09.
- Engineering at Meta, "Meta Andromeda: Powering the Next Generation of Ads Retrieval..." (2026-10-28): https://engineering.fb.com/2026/10/28/ml-applications/meta-andromeda-next-gen-ads-retrieval-gpu-co-design/, checked 2026-11-09.
- Meta for Business, "AI Innovation in Meta's Ads Ranking Driving Advertiser Performance" (2025-03-27): https://www.facebook.com/business/news/ai-innovation-in-metas-ads-ranking-driving-advertiser-performance, checked 2026-11-09.
- Meta for Business, "Three steps to optimize your performance with creative diversification" (2024-11-19): https://www.facebook.com/business/news/three-steps-to-optimize-your-performance-with-creative-diversification, checked 2026-11-09.
- Meta for Business, "The creative advantage: unlocking the power of diversification with Meta Andromeda" (2025-04-22): https://www.facebook.com/business/news/the-creative-advantage-unlocking-the-power-of-diversification-with-meta-andromeda, checked 2026-11-09.
- jonloomer.com, "Meta Andromeda creative diversification" (2025-10-20): https://www.jonloomer.com/meta-andromeda-creative-diversification/, checked 2026-11-09.
- Meta, "Expanding GenAI Transparency for Meta's Ads Products" (updated 2026-06-01): https://about.fb.com/news/2025/02/gen-ai-transparency-metas-ads-products/, 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.