Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.melius.com/llms.txt

Use this file to discover all available pages before exploring further.

When one ad concept is working, the leverage is in scaling it — different products, different angles, different copy, different audiences — without rebuilding from scratch each time. Melius is built for this. A whole variant cascade that used to take a creative team a week typically lands in an afternoon. Time: 30 minutes for a basic 5–10 variant pass; an afternoon for 50+ variants with multiple dimensions. You’ll need: the winning ad (or its constituent parts on a canvas), and a clear list of what you want to vary.

What you can vary

Variant dimensions worth running in parallel:
  • Product / SKU. Same ad, different items from the catalog. Common for retail and DTC.
  • Audience segment. Same product, different model, setting, or lifestyle context.
  • Region. Same ad, localized for different markets — different copy language, different cultural cues, different units. See Localize for international markets.
  • Channel. Same image, every aspect ratio you ship to. See Resize for every channel.
  • Copy / headline. Same image, different CTA or headline.
  • Color / mood. Same composition, different palette — useful for seasonal campaigns.
Pick one or two dimensions to vary at a time. Varying all six at once gives you a matrix of 100+ ads with no signal about which dimension is moving the needle.

Three ways to generate variants

Option 1: Ask the agent for N versions

Best for casting a wide net early. You’ll get diversity but less control over the specific variations.
1

Drop the winning ad onto a canvas

Drag it in from your downloads, or copy it across from another canvas. This is the source the agent will riff off.
2

Brief the agent

The reliable prompt pattern, lifted from how customers actually phrase it:
Help me create 8 variations of this ad, using different creative angles
and briefs. Keep the style and vibe similar otherwise.
Use GPT Image 2 at 2K resolution. Mix up different aspect ratios.
Tweak the count, model, and aspect ratio rules to match the brief. Specifying the model is worth doing — if the variants need legible copy, call out GPT Image 2; if they need style adherence, call out Nano Banana Pro.
3

Let it run

The agent will reason through different creative angles, write a brief for each, and create one image node per variant — all connected to the same source ad and your brand anchor.
4

Review and tweak

Some will land first time. For the ones that didn’t, edit the prompt on that specific node and re-run.
This is the workflow most marketers reach for when they have one ad that’s performing and want to test angle variations. The agent’s strength is in the creative angle dimension — it’ll come up with different framings, scenarios, and copy hooks that a single human brief wouldn’t have produced. You’re outsourcing the angle generation, not the taste call.

Option 2: Multi-variation on one node

Best when you want the same prompt run multiple times to get options — controlling for the probabilistic nature of the models rather than producing genuinely different ads.
1

On the image node, set variations to 3–10

The variations setting is in the node’s right-side panel. Up to 10 in parallel.
2

Hit run

All variations run in parallel via our temporal workflow infrastructure — you don’t wait for one to finish before the next starts.
3

Pick the strongest

Each variation is saved on the same node. Click through them to compare. The unselected ones don’t disappear; you can come back to them.

Option 3: Duplicate the workflow

Best for systematic variation across a known set — every product, every region, every seasonal moment.
1

Build the workflow once

Reference image → style analysis → output image. Get it working for one variant.
2

Select the whole flow and duplicate

Select all the connected nodes (Cmd+A and trim, or drag-select), then Cmd+C / Cmd+V to duplicate. You now have two parallel copies.
3

In the duplicate, swap the input

Use the upload node to swap in the next product photo, or change the reference link, or update the prompt. The downstream nodes will re-run automatically with the new input.
4

Repeat for every variant

Same logic — paste, swap input, run.
Option 3 is the most production-ready. You end up with a canvas where every column is a different variant and every row is the same step in the workflow — easy to scan, easy to hand off, easy to revisit.

Brand-checking variants

The risk of generating variants at volume is that some will drift off-brand. Two ways to keep this in check:
  • Connect the brand anchor to every output node. Every image node should pull from your brand anchor text node as context. If one variant drifts, the brand anchor is your reference for what to fix.
  • Run a final brand-check pass. Once you have your variant set, brief the agent: “Review every variant on this canvas against @brand-anchor. Flag any that drift on palette, typography, or composition, and regenerate them.” The agent will do the audit and re-run the offending nodes.

When to use a scheduled run

If you produce variants on a regular cadence — weekly product spotlights, monthly campaign drops, etc. — you can have an external agent (Claude with MCP, for example) kick off the canvas on a schedule. The pattern most teams use: Claude takes the latest brief on Monday morning, builds out the canvas with all the variants, and pings the designer in Slack with a link when the run finishes. The designer picks the winners. Production capacity that used to require an in-house team is now an overnight cron job. See Drive Melius from Claude.
When generating 50+ variants, keep all output nodes at 1K resolution. Once you’ve picked winners, regenerate just those at 2K or 4K — there’s no reason to render the losers at full resolution.

Common pitfalls

  • Varying everything at once. Diluted signal. Vary one or two dimensions per batch.
  • Not connecting the brand anchor. Variants drift faster the more of them you produce. The brand anchor is the seatbelt.
  • Rendering at 4K from the start. Slow, and most variants will be losers. Render at 1K, pick winners, upgrade resolution selectively.
  • Manual variant creation. If you’re recreating the same setup five times by hand, you missed the “duplicate the workflow” pattern. Build once, paste many.
Last modified on May 18, 2026