Correction, 2026-10-01: the first version of this page priced Nano Banana Pro at $0.04 to $0.08 an image in about 8 seconds and Kling v3 at $0.57 to $0.86 for a 1080p clip; on 8frame Nano Banana Pro is 21 credits an image (median 37 seconds) and Kling v3 Standard returns 720p at 57 credits silent for 5 seconds, so the "under $10" library is about 1,550 credits (roughly $15). It described "batch nodes", which 8frame does not have. It attributed Meta's 22% ROAS figure to Advantage+ creative versus manual settings, when Meta reported it for advertisers turning on Advantage+ creative's AI-driven targeting features. And it stated an assembly time (under 100 milliseconds), a 24-to-48-hour weighting change at Meta, and traditional shoot costs that we could not source; those are removed.
Dynamic creative optimization used to have a supply problem. The engine assembles ads on the fly from a library of interchangeable parts, but somebody had to produce all those parts, and a shoot only gives you a handful of images and one or two headlines. The library stayed shallow, so the "thousands of combinations" pitch stayed theoretical. AI generation fixes the supply side. You can stock a DCO element library with 40 backgrounds, a dozen hero shots and eight motion variants for about 1,550 credits on 8frame (roughly $15 at pack rate), then let the platform's machine learning do what it's good at: finding the winning combination per impression. What you end up with is a genuinely deep library feeding a real optimization engine, built in an afternoon instead of a production quarter.
What DCO actually does
Before the workflow, be precise about the mechanism, because "DCO" gets used loosely. Dynamic creative optimization assembles a personalized ad in real time by combining modular elements, headlines, images, CTAs, backgrounds, product shots, from a library, choosing the combination predicted to work best for each viewer in each context. The assembly happens at serve time, before the impression renders. You are not building one ad. You are building a system of parts and letting the engine build the ads.
This is different from discrete variant testing, where you produce whole finished creatives and split-test them against each other. In DCO, the elements are the unit. The engine mixes them. That distinction drives when you'd choose one over the other, which we cover below.
The 4-step AI DCO workflow
Step 1: Structure the element library before you generate anything
DCO only works if your elements are truly modular, meaning any headline can sit on any background under any CTA without looking broken. So design the slots first. A typical Meta or programmatic setup needs: backgrounds or scene beds, hero product shots, motion clips (for video placements), headlines, and CTAs. Decide how many of each slot you want before generating a single asset. Ten backgrounds by six heroes by four CTAs is 240 combinations from 20 produced elements. That multiplication is the entire point.
Step 2: Generate the visual elements with Nano Banana Pro and Kling v3
Use Nano Banana Pro for the still elements, backgrounds, product hero shots, texture and detail frames. On 8frame it is 21 credits an image at 1K or 2K (42 at 4K), with a median generation time of 37 seconds. Pick a single product reference image and attach it to every generation so the product stays as close as possible across every background, then check each output against it. That consistency is non-negotiable in DCO: the engine will pair your product with dozens of scenes, and if the product drifts between generations the whole library looks incoherent. Reject drifted frames rather than shipping them.
For placements that take video, generate motion beds with Kling v3 Standard: 57 credits for a silent 5-second clip (86 with sound), at 720p, with a median generation time of about 60 seconds. If the placement needs 1080p, Kling 2.6 Pro is 47 credits for a silent 5-second clip. Keep these motion elements neutral enough to pair with any headline overlay. On 8frame you build this on one canvas: the product reference image connected into a Nano Banana node for stills and a Kling node for motion, using a chosen still as the start frame. There is no batch node, so each element is its own run: duplicate the node, change the prompt, run it again. (Nano Banana Pro returns one image per run; Nano Banana 2 can return up to four.)
Step 3: Write the text elements to combine, not to stand alone
The copy slots are where most DCO libraries fail. Headlines and CTAs have to read cleanly in any combination, so avoid writing a headline that only makes sense over one specific image. Write four to eight headlines that each carry the value claim on their own, plus three to four CTAs across urgency, benefit, and neutral framings. Keep them short enough to overlay on any of your generated backgrounds without crowding the composition.
Step 4: Load the library into the platform engine and let it assemble
For Meta, this means Advantage+ creative. You supply the assets and the objective, and the system varies and assembles them. Meta's description of the standard enhancements includes adjusting image brightness and contrast, varying the image aspect ratio, adding templates so images fit feed placements, and repositioning the text you supply as primary text, headline or description based on expected engagement. On results, Meta's engineering team wrote in December 2024 that advertisers who had not used Advantage+ creative and turned on its AI-driven targeting features saw a 22% increase in ROAS from Meta ads. That is Meta's own figure about its own product, so treat it as directional rather than a promise. For programmatic DCO outside Meta, you upload the modular elements to a DCO-capable ad platform that runs the real-time assembly across the open web.
Load your full library at launch rather than trickling elements in, so the engine has combinations to test from the first day of delivery.
Cost math: AI library vs. traditional DCO shoot
A traditional DCO library sourced from a shoot means paying for a production day, then a retoucher to cut out enough element variations to make assembly worthwhile, and every refresh means another shoot. Shoot costs vary too much by market and brief to put one number on, so the table shows what the AI side costs on 8frame in credits (a credit is about $0.01 at pack rate).
| Line item | Traditional route | AI on 8frame |
|---|---|---|
| 40 background/scene stills | Shoot day plus retouching | 40 x 21 = 840 credits (Nano Banana Pro, 1K or 2K) |
| 12 hero product shots | Part of the shoot day | 12 x 21 = 252 credits |
| 8 motion clips (5 s, silent) | Motion shoot or edit | 8 x 57 = 456 credits (Kling v3 Standard, 720p) |
| Total | A production budget | 1,548 credits, about $15.48 |
Rejected takes cost the same as kept ones, so budget for some re-runs. The AI library also refreshes: when elements fatigue, you regenerate the tired slots (21 credits a still, 57 a silent clip) instead of booking another shoot, which matters because DCO libraries fatigue the same way discrete creatives do.
When DCO beats discrete variants (and when it doesn't)
Use DCO when you have many audience segments and contexts, and the winning message genuinely differs by viewer. Personalization at the element level is DCO's home turf: retargeting with dynamic product feeds, multi-market campaigns, large prospecting audiences where the engine can find per-segment combinations you'd never guess.
Use discrete variant testing when you're still hunting for the concept itself. DCO optimizes within a system of parts, but it can't tell you that your whole angle is wrong. Find your winning concepts with AI ad variant testing first, then feed the proven elements into DCO to scale the personalization.
The measurement trap to avoid: DCO reports a blended result across thousands of combinations, so a strong average can hide a few dominant combinations doing all the work while the rest drag. Check the combination-level breakdown, not just the campaign average, or you'll keep paying to assemble elements that never win.
Related reads
For the underlying mechanism explained from scratch, see what is dynamic creative optimization. For why a deep element library is a fatigue defense as much as a personalization tool, see creative fatigue and the AI refresh math.
FAQ
What does it cost to build a DCO element library with AI?
On 8frame, 40 background stills and 12 hero shots on Nano Banana Pro (21 credits each) plus eight silent 5-second Kling v3 Standard motion beds (57 each) come to 1,548 credits, about $15.48 at pack rate, before any re-runs.
Does Advantage+ creative replace a DCO platform?
On Meta, Advantage+ creative does the assembly and variation for you: it can adjust brightness and contrast, vary aspect ratios, add templates and reposition your text. Outside Meta, you need a DCO-capable ad platform to assemble the elements across the open web. Either way, the engine can only choose among the elements you give it.
When should I use discrete variants instead of DCO?
When you are still looking for the concept. DCO optimizes within a system of parts but cannot tell you the angle itself is wrong. Find the winning concept with discrete variant tests first, then feed its elements into DCO to personalize at scale.
Sources
- Meta Andromeda and the 22% ROAS figure: Meta Engineering, December 2, 2024, checked 2026-10-01
- Advantage+ creative standard enhancements: Meta for Developers, April 14, 2023, checked 2026-10-01
- 8frame Nano Banana Pro, Nano Banana 2, Kling v3 Standard and Kling 2.6 Pro prices and outputs: the 8frame tool registry, 2026-10-01
- Generation times: 8frame production data, read-only, medians over 120 days to 2026-09-30
Build your DCO element library on one canvas. Attach a product reference, generate 40 backgrounds and 8 motion beds on 8frame, and hand the engine a library deep enough to actually optimize.
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.