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.

This is the most-run workflow on Melius. You have a look you want to hit — your own past campaign, a competitor’s ads, a Pinterest board, a fashion editorial — and you want a fresh batch of statics in that style. Time: 10–15 minutes from open canvas to first downloadable assets. You’ll need: 2–10 reference images, a clear sense of the product or subject, and a brand anchor if you have one (how to make one).

​
The workflow

1

Open a new canvas

Name it for the campaign. If you have a brand anchor, paste it in as a text node now and keep it parked at the top of the canvas — you’ll connect things to it later.
2

Drop your references in

Drag 2–10 mood images onto the canvas. Two to four is usually plenty; more than ten often hurts.
3

Group them as a unified group

Select them all, right-click → “Unified group”. This gives you one output port for the whole set.
4

Add the Image Style Analysis template

Templates menu → “Image style analysis”. A pre-configured text node appears on the canvas, set up to use Gemini 3.1 Pro (the best model for this task).
5

Connect references → style analysis

Drag from the unified group’s output into the style analysis node’s reference input. Hit run on the text node.
6

Read what came out

The style analysis output is a detailed description: palette, lighting, mood, composition, depth of field, shot type. If anything seems off, edit it directly — the output is just text, you can rewrite it.
7

Create your output image node

Right-click on the canvas → New image node. Pick Nano Banana Pro (best for style adherence) at 2K resolution and the aspect ratio you need (9x16 for stories, 4x5 for feed, 1x1 for legacy, 21x9 for email banners).
8

Connect the style analysis output to the image node

Drag from the style analysis text output into the image node’s input.
9

Write the prompt

Keep it short. The style description is doing the heavy lifting. Example:
Generate a hero image of [product] [in this scenario].
Use the style description below as the visual reference.
10

Run 3–4 variations

Set the variations count to 3 or 4 before hitting run. The model is probabilistic — running multiple in parallel is faster than running one, getting a bad result, and trying again.
11

Pick winners, tweak the rest

Some variants will land, some won’t. For the close-but-not-quite ones, edit the prompt directly on that image node — “keep everything else the same, but…” — and re-run. Each re-run is saved as a new version on the same node, so you can compare without losing the original.

​
A worked example

Brand: boutique apparel. Goal: five product hero shots in the style of a fashion editorial moodboard. The moodboard (six images, all painterly product photography with warm tones) goes into a unified group. Style analysis runs on the group and outputs:
“Editorial product photography. Warm golden-hour lighting from camera-left. Shallow depth of field, painterly bokeh. Product centered with negative space above. Color palette: cream, terracotta, soft black. Composition feels like a 1990s magazine layout — generous margins, no graphic overlay…”
The image node is set to Nano Banana Pro, 4x5, 2K. The prompt: 
Hero shot of @hoodie in the style described below.
The hoodie is the subject. Soft natural light. Editorial composition.
Use the style description for color, lighting, and mood.
Run with 4 variations. Two are usable as-is, two need minor tweaks (one has the hoodie too small in frame; the other has color that’s drifted toward orange instead of terracotta). Both fixes happen on the node — edit the prompt, add the constraint, re-run. Total time from blank canvas to four production-ready hero shots: about 12 minutes.

​
Common pitfalls

  • Forgetting to run the style analysis node. If you connect everything and skip the run step on the text node, the analysis output is empty and the image node has no style guidance. Always run the text node before the image node.
  • Mixing incompatible references. Five fashion editorials and one screenshot of an ecommerce storefront will confuse the style analysis. Keep the moodboard internally consistent.
  • Picking the wrong model. Nano Banana Pro is best for style adherence. GPT Image 2 is better when the output has a lot of legible typography. Use the right one for the job — see Pick the right model.
  • Skipping the unified group. You can connect each reference image to the style analysis node individually, but it’s tedious and easy to forget one. Unified groups handle this in a single edge.
Once you have a style analysis text node that’s working well, you can copy and paste it across canvases — same brand, different campaigns. It’s a small reusable artifact that pays off every time.
Last modified on May 18, 2026