Generate On-Brand Marketing and Social Creatives at Scale

Most marketers who need a batch of social visuals hit the same wall: a busy designer, a Canva template that looks like every competitor, or stock that looks like every competitor. You can generate a consistent set of on-brand creatives straight from your agent, using a reference image to anchor your brand and gpt-image-2's text rendering to put real headlines in the frame.

The workflow: pick a brand reference, write a style prompt, and generate a batch across formats that all read as one family. Plus where it breaks down, because the honest answer is "most of it, not quite all."

Why agents fit social creative generation

The agent path: describe what you want, pass a reference for brand continuity, get a usable creative in under thirty seconds. That is a full week of posts in one session, or a stack of ad variants before a campaign. AgentBrush is the MCP server that makes it work inside Claude or Cursor: plain-language request, file lands in your project, no API key in your shell.

Step 1: Establish your brand reference

The highest-leverage move for consistency is starting with a reference. Pass your logo, a hero illustration, or a product shot, and AgentBrush sends it alongside every prompt as a visual anchor for color, shape, and aesthetic. No brand visual yet? Generate one and save it first.

Generate a clean brand visual: a stylized circuit-board robot mascot with a minimal geometric
design, two accent colors (electric blue and slate gray), isolated on a white background.
This will serve as the visual identity reference for all marketing assets.
preset: flat_illustration · quality: high · 1:1

Save the robot image as a reference named "brand-mascot".

Meet the mascot. He is about to spend this entire article selling optimism he does not appear to feel. If the expression looks familiar, it is Marvin from The Hitchhiker's Guide to the Galaxy: brain the size of a planet, reassigned to social media. There is a real point under the joke. A reference locks in expression along with color and shape, so whatever face your mascot pulls in the first render, you are shipping it across the whole campaign. Every later generation can now call brand-mascot by name.

Step 2: Generate a social post with a real headline

This is where gpt-image-2's text rendering earns its keep: it renders the exact strings you give it instead of guessing at letters. Put the copy in the prompt and it appears.

Using the brand-mascot reference, generate a social post creative (4:5 portrait for Instagram).
The robot is centered, friendly pose, gesturing at a bold headline below it.

Exact text to include:
- Large headline: "Ship Faster. Look Good Doing It."
- Subtext below: "AI-generated visuals, inside your agent"
- Small footer: "agentbrush.dev"

Electric blue and slate gray palette, off-white background, clean flat illustration style,
generous negative space. No generic stock art feel.
preset: flat_illustration · quality: high · 4:5
reference_image_paths: ["brand-mascot"]

He is gesturing at "Ship Faster. Look Good Doing It." with the body language of a robot who has read the roadmap. The takeaway survives the joke, though: put the exact strings in the prompt and the model renders them verbatim. It does not improvise copy, so treat the text fields like a brief, not a hint.

Step 3: Extend to a batch of formats

Once the first asset works, extend it. Pass the same reference, keep the same style descriptors, and vary only the aspect ratio and context.

Square (1:1) for feed posts:

Using the brand-mascot reference, generate a square (1:1) version of the social creative.
Same palette, same robot, same headline copy. Tighten the composition for a square crop.
preset: flat_illustration · quality: high · 1:1
reference_image_paths: ["brand-mascot"]

Landscape (16:9) for an X banner or slide header:

Using the brand-mascot reference, generate a landscape banner (16:9).
The robot stands on the left third, large headline on the right: "AI Creatives. No Designer Needed."
Electric blue and slate gray, flat illustration, plenty of breathing room.
preset: flat_illustration · quality: high · 16:9
reference_image_paths: ["brand-mascot"]

Ad creative, a different angle (3:1 wide hero):

Using the brand-mascot reference, generate an ad creative for a landing page hero (3:1 wide).
Same robot, this time pointing at copy beside it: "Generate. Stay On-Brand. Ship."
Clean background, high contrast between the text and the background for readability.
preset: flat_illustration · quality: high · 3:1
reference_image_paths: ["brand-mascot"]

"AI Creatives. No Designer Needed." He does not look reassured by that either. But he carries cleanly across every format, because each call passes the same reference, and a coherent set drops in a few minutes.

Style hints that hold a batch together

Consistency comes from two places: the reference (appearance) and the style descriptors in your prompt (rendering). Keep the descriptors identical across the batch.

  • preset: pick one and hold it. flat_illustration for vector-feel, realistic for lifestyle or product.
  • Palette: name the colors. "Electric blue and slate gray" beats "our brand colors."
  • Negative-space cues: say what you do not want. "No generic stock art feel," "no texture or shadow."
  • Typography: describe weight and case if you care. "Heavy sans-serif, all-caps headline."

Token cost for a batch

Each high-quality image is 20 tokens at 1024x1024, so a batch of ten is 200. That is a fraction of Pro (600 tokens, $14.99/month) and sits well inside Power (1,300 tokens, $29.99/month, then $0.04/token overage). Draft at medium (5 tokens) for layout or low (1 token) for direction before the high-quality run. A weekly five-to-ten-image calendar fits Pro comfortably; a launch-week variant push wants Power.

Where it breaks down

Exact brand fonts. The model matches a typeface to your description; it does not load your font file. For a licensed typeface, take the generated layout into a design tool and swap the real font for finals.

Complex multi-column layouts. One or two content areas is its comfort zone. Dense ad layouts with many CTAs, fine print, or a product grid belong elsewhere. Use it for the visual layer.

Exact hex color. "Electric blue" lands plausibly blue, not guaranteed #0066FF. Post-process if compliance matters.

Your actual product. A reference helps, but for a specific SKU with exact details, pass the product photo and describe the scene around it. The model respects shape and color; it is not a photographic composite.

None of these are reasons to skip the workflow. They are reasons to use it as the fast draft layer and bring precision tooling for the final 10%.

Pair it with character consistency

If your brand has a mascot, sad or otherwise, the reference does double duty: consistent style and consistent character, with no training. Generate the canonical mascot once, save it, and every post uses it correctly. For abstract brand visuals and landing-page illustration sets, see AI flat illustrations; for working from an existing brand photo, generating consistent images from references.

FAQ

Can I use this to generate ad creatives with real text in them? Yes, and it is one of the stronger use cases. Provide the exact strings, specify the typographic direction (weight, case, placement), and gpt-image-2 renders them accurately. Treat the text fields like copy directions, not suggestions.

How do I keep a batch of images looking like a family? A shared reference image plus identical style descriptors on every call. Same reference_image_paths, same preset, same palette and style terms. The consistency lives in the prompt, not just the model.

Does this work for generating AI social media images in different languages? Yes. gpt-image-2 has strong multilingual text rendering. Specify the language and the copy and it renders the correct script. Test right-to-left scripts in particular, since layout direction matters.

What is the difference between a preset and a style hint? The preset (flat_illustration, realistic) sets the rendering mode; style hints ("generous negative space," "hard-edged shapes") fine-tune within it. The preset is the base, the hints are the adjustment.

How many images can I realistically generate in one session? As many as your token balance covers; the real limit is your review time, not the ceiling. A launch-week set (cover, three social variants, two ad crops) is about 6 high-quality images, 120 tokens, inside a Pro plan.

The fastest test is one prompt. Connect AgentBrush to your agent, pick a brand reference you already have, and generate a single creative. If it is usable, the batch follows. Your mascot does not have to enjoy it. Ours certainly does not.