To generate ad creative variations with AI, do not ask an agent for “twenty fresh ideas.” Give it one invariant campaign brief and a grid of controlled changes. Let it coordinate model discovery, accepted inputs, price checks, generation jobs, names, and review status. Keep the offer, product facts, policy judgment, spend approval, and final selection human.
The goal is not maximum output. It is a set in which every image has a reason to exist.
| Keep fixed | Change deliberately | Never let the agent infer | |---|---|---| | Product and SKU | Visual hook | Product claims | | Audience | Setting or context | Offer terms and dates | | Offer and proof | Composition | Usage rights | | Placement and crop | Product angle | Policy approval | | Measurement plan | Color or lighting family | What a winner means |
Build current image concepts in Oakgen's AI image generator, compare model directions in the Image Arena, or produce UGC-style work in UGC Ads. The connected-agent pattern below is preparation for a later image-only MCP release.
The public MCP is not in production. This guide does not claim a connection endpoint, credential flow, named-host result, plan requirement, or launch date. The intended first surface contains image operations only.
Who this workflow is for
This is for performance marketers, creative strategists, DTC teams, and agencies that need more testable still-image directions without losing the campaign's logic.
It assumes somebody owns the media plan. An image agent cannot decide whether an apparent winner is incremental, whether the test received fair delivery, or whether a claim is permitted. It can make the production system easier to operate.
If you run client work, start with the wider creative MCP workflow for marketing agencies. If model choice and cost are the main problem, use the image-model selection and pricing worksheet.
If the connected service itself is still undecided, compare candidates in the AI image generation MCP buyer's guide before you standardize the batch.
Research note
This guide was reviewed on July 24, 2026 against Meta's official creative-diversification, photo-ad, and ad-review guidance; the official Model Context Protocol introduction; and Oakgen's deterministic connected-generation contract. It does not claim a performance lift or present generated examples as tested ads.
Meta's own guidance encourages creative variety and experimentation, but “more variants” is not a measurement plan. Platform tools may also adapt creative for placements or audiences. Record what you supplied, what the platform changed, and what the test is actually comparing.
Write the invariant brief first
Copy this into the campaign record before prompts:
| Brief field | Approved value | |---|---| | Campaign job | What commercial action should this support? | | Audience | Which specific buyer and moment? | | Product | Exact SKU, package, color, and included parts | | Offer | Approved language, value, dates, and exclusions | | Proof | Facts the brand can substantiate | | Required marks | Logo, legal copy, or later layout zones | | Prohibited territory | Claims, depictions, props, and contexts to avoid | | Placement | Ratio, safe area, resolution, and destination | | Test variable | The one major creative dimension changing | | Success measure | Metric, observation window, and decision owner | | No-ship conditions | Defects that trigger rejection |
The agent may point out that a field is missing. It may not fill a commercial fact from context.
The quote-before-run agent workflow
1. Turn the testing question into cells
“Does visible product use beat a static hero composition for first-time buyers?” can become three execution cells per hook. “Find something that converts” cannot.
Ask the agent to return IDs, hypotheses, and changed variables before any image prompt. This is where a creative strategist removes duplicate or untestable ideas.
2. Search models by job
Oakgen's intended workflow uses search_models to find image models suited to the brief. The agent should explain why an option fits: reference preservation, layout direction, text treatment, speed for concepts, or another job-relevant property.
The newest model is not automatically the right one. Neither is the cheapest.
3. Inspect the schema
get_model_schema is intended to expose the selected model's actual accepted inputs. Check required fields, supported reference inputs, output shape, and any count or sizing controls. If the model cannot express the plan, change models or change the plan.
Here is where this breaks: an agent can write a beautifully specific instruction for a control the model does not have. The schema is more authoritative than the agent's confidence.
4. Quote a proof set
Use get_pricing on the exact request. Select one hard representative from each creative hook, then get human approval for its model, payload, output count, and quoted amount.
Do not price one image and treat that as approval for a matrix. Expand only after the proof set passes product, brand, and composition review.
5. Start every approved request once
start_generation is intended to begin durable work. Give each request a stable identity derived from the campaign, cell, version, and approved payload. In plain language, this is the order number for the generation. A retry with the same order number should not become an excuse to purchase a duplicate.
Save the returned job handle. Use get_generation_status when the work is slow or the chat is resumed. Do not send a second start just because a response disappeared.
6. Review before expansion
Sort the result into approve, revise, or reject:
- Approve means the hypothesis is visible and invariants hold.
- Revise means the direction works but one named defect is fixable.
- Reject means the product, claim, brand, policy, or hypothesis is wrong.
Only approved directions receive more executions.
Build a proof set, not a generation dump
Compare a few deliberate image directions in Oakgen before you expand the campaign into a larger testing matrix.
The ad-variation review grid
This is the article's linkable asset. Use one row per image and require a written reason for every decision.
| Field | Example | Reviewer question |
|---|---|---|
| Asset ID | summer_a02_feed45_v01 | Can this result be traced to one approved request? |
| Hypothesis | Product-in-use makes the benefit concrete | Is the hypothesis visible without reading the prompt? |
| Changed variable | Setting: cramped apartment kitchen | Did anything else substantial change? |
| Product accuracy | Pass / revise / reject | Is this the actual SKU with real components? |
| Offer and claim | Pass / revise / reject | Does the visual imply more than the approved brief? |
| Brand fit | Pass / revise / reject | Is the direction recognizable without relying on a logo? |
| Placement | Pass / revise / reject | Does the focal point survive the final crop? |
| Policy review | Pending / pass / reject | Has the publishing owner checked current rules? |
| Decision | Proof, expand, revise, stop | What happens next, and who owns it? |
Use a stable name:
campaign_hypothesis-cell_placement_version_status.ext
Do not label five nearly identical renders “v1–v5” without recording the prompt seed, input, or meaningful execution difference. A version number is not provenance.
For current production, generate the four proof cells in Oakgen, then place the outputs in this grid before deciding which hook deserves more executions.
Five prompts for controlled creative work
1. Generate the plan, not the ads
Using the invariant brief below, propose four visual hypotheses with three executions each. Change one major creative variable per hypothesis. State what every cell tests and flag any brief field that is missing. Stop before model search or generation.
2. Protect the product reference
Treat the uploaded product images as authority. Preserve SKU, geometry, color, logo placement, package text, and included components. If the selected image model cannot accept the required reference input, stop and recommend a different model rather than approximating the product.
3. Enforce quote approval
For the four approved proof cells, search for suitable image models, inspect the accepted inputs, and quote each exact request. Return a compact approval table. Do not start paid work until I approve every row.
4. Start safely
Start only the approved cells using their stable asset IDs. Return one job handle per cell. If a response is uncertain, retrieve status using the same handle or stable request identity; do not create another start.
5. Review without flattering the output
Evaluate each completed image against the invariant brief and its named hypothesis. Mark product accuracy, claim safety, brand fit, crop safety, and visible artifacts as pass, revise, or reject. Do not call a result successful because it is polished.
A practical creative testing framework
Exploration phase
Use materially different hooks to identify promising territory. Multi-variable change is acceptable here if the team labels it as exploration rather than a clean test.
Controlled test phase
Choose the best directions and isolate the variable. Keep audience, offer, placement, media setup, and evaluation window as stable as the actual platform permits.
Expansion phase
Build more executions only after a direction earns attention. Vary composition, setting, or product angle deliberately. Do not multiply a weak idea to satisfy a content quota.
Learning record
Capture what the result suggests and what it does not prove. “Contextual use beat a plain hero in this test” is useful. “People hate hero images” is not.
Meta notes that its systems may generate or adapt variations through Advantage+ creative. If those features are enabled, record them. Otherwise, the media team can mistakenly attribute a platform-generated change to the source asset.
Failure handling
A price is higher than expected
Compare the payload with the proof quote. Check output count, dimensions, model route, and reference inputs. Reduce scope or request a new quote; do not silently spend through the difference.
A start response is lost
Use the stable request identity and existing handle to recover status. The safe rule is check before starting again.
A job remains non-terminal
Keep the handle and follow the MCP troubleshooting sequence. Decide whether to wait or request cancellation. A cancellation request still needs final-state confirmation.
Product details drift across variants
Stop expansion. Improve the reference set or choose a workflow better suited to product preservation. Prompt repetition is not quality control.
Generated copy is misspelled
Do not keep rolling the image model until it gets lucky. Generate or preserve the visual, then add exact typography in a controlled design step.
Common mistakes
- Choosing a model before defining the job.
- Calling twenty unrelated images a test.
- Changing hook, copy, color, crop, and audience in the same formal test.
- Approving at thumbnail size.
- Treating platform acceptance as legal or brand approval.
- Restarting a slow job instead of checking status.
- Expanding a direction before the proof set preserves the product.
- Reporting a winner without recording platform-side creative changes.
Frequently Asked Questions
How do I generate ad creative variations with AI?
Lock the product, offer, audience, claim, placement, and measurement plan first. Then vary one meaningful creative dimension, review a small proof set, and expand only the directions worth testing.
What is an invariant brief?
It is the approved part of the campaign that every variation inherits: product truth, audience, offer, proof, mandatory branding, prohibited claims, placement, and the test's success measure.
How many variables should change in one ad test?
Change one major variable when you need a clean learning, such as visual hook or setting. Multi-variable concepts can explore direction, but they are harder to diagnose as formal tests.
Why should an AI agent check the image model schema?
The schema shows which inputs the selected model actually accepts. It prevents the agent from inventing controls or choosing a model that cannot use the required reference or output shape.
How do I avoid duplicate generations?
Give each approved request a stable identity, save the returned job handle, and retry status checks rather than starting again when a response is slow.
Is Oakgen's image-generation MCP available now?
No. Public production access is not released. The planned six-operation image surface is described here as a future workflow, without an endpoint, credential process, host promise, or launch date.
Can I build ad creatives in Oakgen today?
Yes. Oakgen's current web tools include image generation, image-model comparison, agent chat, and UGC ad workflows. These current routes are separate from the unreleased public MCP.
Sources and Further Reading
- Meta: expand your ad creative strategy — official guidance on placements, formats, and creative diversification.
- Meta Advantage+ creative — official explanation of platform-side creative generation and adaptation.
- Meta photo ad guidance — official creative guidance for focal point, consistency, resolution, and preview.
- Meta ad review guidance — official overview of the review system.
- Model Context Protocol introduction — official protocol roles and tool concepts.
Start with one deliberate proof set. Generate the four hardest image concepts in Oakgen, review them against the grid, and earn the right to batch.
