How to build a static ad from a style reference on Melius

Key takeaways

• You brief an agent; you don't wire the workflow. One instruction naming your references and the output you want is enough. The agent picks the models, writes the prompts, creates the nodes and connects them up.
• The steps still happen, and you can watch them happen. Reference grouping, style analysis, model choice, variation count: the agent does each one as a visible node rather than behind a loading spinner.
• Expect to finish the last mile yourself. We put the agent's first pass at 85% to 95% of the way there. The rest is prompt tweaks on individual nodes, and that's where your judgement goes.
• Brief the whole job, not the next step. Asking for the references, the variations, every ratio and a brand check in one instruction costs the same as asking for one of them.
• Ambiguity gets a question, not a guess. In ask-permission mode the agent checks aspect ratio, audience and tone before it spends a generation. In auto-run it decides and builds while you're elsewhere.

Static ads that match a look you already have

To build a static ad from a style reference on Melius, you brief the Mel agent: point it at your reference images and ask for new statics in that style. It assembles the canvas, runs the style analysis, chooses the models, generates the variations, and leaves every step visible so you can correct the ones that missed. From brief to assets you can download runs about 10 to 15 minutes.

Most of the work in matching a look goes into describing it. You know the campaign you want to sit next to, and you can see exactly why it works, but turning that into instructions a model will follow is the step that eats an afternoon. That's the step you're handing over.

What you actually type

Right-click anywhere on the canvas and choose Ask Mel, or press the forward slash key. Anything you have selected goes in as a reference automatically. Then describe the job. You don't need to name the model, set the resolution, or say which output connects to which input. Mel will do these for you.

Our standing advice is to ask for more than you think will work. A brief naming the references, the number of variations, every ratio you ship and a brand check against your brand anchor costs no more to write than one asking for a single image, and if the agent can't cover the whole job it tells you. Brief it the way you'd brief a designer, with a concrete goal and real constraints, not the way you'd chat. This is the same brief-to-execution workflow applied to a single look.

What the agent does in the background

None of this needs doing by hand, and all of it stays inspectable while it happens. A cyan marker on the canvas shows where the agent is currently building.

• Reads your brand anchor. If you keep one, the node holding your voice, palette and type gets pulled in as context before anything generates.
• Groups the references. Your reference images become a single unified group, so the analysis reaches all of them on one edge instead of one per image.
• Runs the style analysis. The Image Style Analysis template comes preconfigured but has inputs that need wiring, which is exactly the case to hand to the agent. It returns a written description: palette, direction and quality of the light, mood, composition, depth of field, shot type.
• Picks the model per step. Melius recommends Nano Banana Pro for hewing to a photographic look, and GPT Image 2 where the ad carries visible copy, its High tier for production output. You can override that on any node.
• Writes the prompt. Because the style description carries the look, the generation prompt only has to name the subject and where it sits. Short prompts win here, and the agent knows it.
• Builds the image nodes and connects them. Style description into the image node, resolution set, ratio set per placement. The node hides any ratio the chosen model can't produce.
• Runs the variations in parallel. Image models are probabilistic, so several takes at once beats one take and a retry. Ask for every ratio in the same brief and it queues those together too.

Where you step in

The first pass gets you most of the way, not all of it. We put that at 85% to 95%, and the last stretch is the part worth your attention, because it's taste rather than assembly.

Work at the node, not in the chat. Open the one image that nearly landed, edit its prompt with the specific correction, and re-run that node alone. Product too small in frame becomes a line saying the product fills 40% of the frame width. Colour drifting green becomes a warmer grade with terracotta in the highlights. Illegible type on a label means switching that node to GPT Image 2 High.

Go back to the agent for the genuinely bulk things: swapping a model across multiple nodes, regenerating everything with the brand anchor attached, resizing ten nodes into three ratios each.

One habit worth forming. The agent doesn't automatically know about an edit you made by hand, so if you later ask it to repeat that treatment across the canvas it may revert to its original approach. Either make the change on a shared context node like the brand anchor, where it propagates on the next run, or tell the agent what you changed and ask it to apply that too.

Where this workflow goes wrong

• The brief was too small. Describing one step, running it, then describing the next is the safe instinct and it costs you the afternoon. A re-brief is cheap; orchestrating by hand is not.
• The references disagree with each other. Five fashion editorials and one storefront screenshot do not share a look, and the description will average them into something you didn't ask for. The agent can only describe what your set has in common.
• The agent got treated like a chatbot. It isn't a chat window, it builds canvases. Conversational back-and-forth gets far less out of it than one concrete brief carrying references and constraints.
• One node got edited fifteen times. Past about the fifth tweak the problem usually isn't the prompt. Step back and re-brief, because the agent has the whole canvas as context and will often unstick it.
• The output was judged on the canvas preview. Previews are downscaled at every resolution to keep the canvas quick. Open a variation full screen, or download it, before you rule it out.

Why the canvas is still there

Melius is the Agentic OS for creative work, and underneath the agent is a node-based canvas where every model call is a node and edges pass one node's output into the next. The canvas isn't something you have to learn before you can work. It's there so the agent's decisions are legible: which model ran, what prompt it wrote, which reference fed which step.

That's what separates this from describing what you want in a chat and hoping. Nothing here is a black box. When an output is wrong you find the node where it went wrong, change that one thing, and re-run everything downstream of it. Earlier node canvases were built before agents existed and assumed a human would wire up every node; this one exists so you can see and steer the work instead of performing it.

The models sit behind one subscription. Melius spans a library of leading models for image, video and audio, and the agent routes each step to the right one unless you choose per node.

You own the content you create on Melius and can use, publish, and distribute it commercially, subject to the underlying model providers' terms. Melius doesn't claim ownership of your work and doesn't use it to train AI models.

Have a look you want to hit? Open Melius, drop your references on a canvas, and tell the agent what you want.

Frequently asked questions