How to generate a product hero shot on Melius

Key takeaways

• A hero shot needs three separate inputs, not one prompt. A clean product shot, your product's real specifications, and a style reference each go in as their own node, and the prompt's job is mostly to say what must not change.
• Branding is the part the model invents. An image model regenerates a logo rather than placing your logo file, so the mark comes back nearly right and specifically wrong. Give it the artwork as its own input and tell it not to redraw it.
• A lifestyle photo is not a product reference. The model needs a clean, evenly lit view of the product itself. Mood references and product references do different jobs and go in separately.
• Generate at 2K or higher, not 1K. Fine detail is where fidelity is won or lost: fabric weave, embossing, label type. More resolution gives the model more to work with on exactly those features.
• Correct on the node, don't start over. When scale or placement is off, edit that node's prompt with the specific fix and re-run it. Re-prompting the whole canvas throws away the parts that worked.

Product hero shots from a pack shot

To generate a product hero shot on Melius, put a clean pack shot of the product on the canvas, add its real specifications and a style reference as separate inputs, connect all three into an image-to-image node at 2K, and write a prompt that names the scene while stating plainly that the product itself must not change. Then run three or four variations and inspect them full screen.

The workflow is easy. Getting the product to look like your product is the part that takes practice, and it's the complaint marketers raise most often on onboarding calls. The shot is beautiful and the proportions are subtly off, or the logo resembles your logo, or the fabric drapes like a stock photo rather than like the thing you sell. Nearly all of that traces to one cause, and it's fixable.

Why generated products come back nearly right

An image model doesn't paste your product into a scene. Whatever falls inside the region it generates gets rebuilt as an approximation from what it can infer, so anything it can't infer gets invented. Your logo isn't a shape it can reason its way to, real proportions aren't recoverable from one photograph, and the weight and fall of your material isn't something a prompt adjective conveys.

What fixes it is giving the model your actual assets as distinct inputs, then being explicit about which of them are fixed. Prompt wording alone won't get you there. Every step below removes one more thing from the set the model is guessing at.

Building the shot, step by step

The click path is in the help centre recipe for product photos and swaps. The shape of the work is this:

1. Drop your pack shot on the canvas. A clean product photo on a neutral background. If you have several angles, bring them all in and unified-group them, so the agent works from every angle rather than one.
2. Give the agent the real specifications. Ask it in chat to pull dimensions, materials and colour from your product page URL and save them as a text node. It now has your stated specs to work from rather than inferring everything from a photograph.
3. Add the logo as its own node. If the product carries visible branding, put the logo file on the canvas separately and connect it in. In the prompt, tell the model to use that reference exactly and not to redraw it.
4. Add a style reference for the look. A moodboard or a past campaign, turned into a style description. It governs light, composition and mood, and stays separate from everything describing the product.
5. Create an image-to-image node at 2K. Melius recommends Nano Banana Pro for material and lighting fidelity, and GPT Image 2 (High) where packaging text has to be legible (guidance current as of September 2026). Set the ratio you're shipping, and go 2K or 4K rather than 1K: the small features are the ones that give a shot away.
6. Wire the inputs to their jobs. Pack shot into the image input. Specifications, style description and logo into context. Then reference them by name in the prompt with @ mentions so the model knows which input governs what.
7. Write the prompt around what stays fixed. Name the scene in one line, then spend the rest on constraints: use the exact product from the reference, don't restyle it, don't change its colour, don't modify its shape or proportions.
8. Run three or four variations and inspect them full screen. The canvas preview is downscaled, and downscaling hides precisely the defects you're checking for. Open each candidate properly before you pick one.
9. Check the output against the actual item. Shape, colour, logo, any packaging text, proportions. A clean pack shot and real specs improve your odds; they do not remove the review step.

Where product shots go wrong

• A lifestyle photo was used as the product reference. The product is small, partly hidden and lit for mood, so the model never gets a clean look at it. Add a studio shot or a transparent PNG as a separate product input.
• No specifications went in. Dimensions and materials get inferred from a single image, and the model has no way to know how big the thing is or what it's made of.
• The logo was left to the model. Branding is the one element where close enough is still wrong, because anyone who knows the brand knows what the mark should look like. It needs to be an input, not an output.
• The whole thing got re-prompted over one flaw. When the product is 20% too small or facing the wrong way, that's a one-clause correction on the existing node, not a fresh start.
• Typography and material were asked of one model. Rendering legible packaging copy and rendering convincing material are different problems, and a canvas can use a different model at each step. Our per-job model guidance is in the help centre, and it changes as models ship.

Scaling one shot across a catalogue

Once a canvas produces a shot you'd ship, the graph is the asset. Duplicate it, swap the pack shot and the specifications for the next SKU, and the light, composition and constraints carry over. That's how a team working through a catalogue gets consistency instead of forty shots that each look slightly different.

The neighbouring jobs start from the same graph: background variants for seasonal scenes, a swap to drop your product into a scene you already have, re-composition for each channel's ratio. Melius ships a Product Swapper preset for the swap once it becomes routine.

Why this runs the way it does on Melius

Melius is the Agentic OS for creative work, built on a node-based canvas where each input and output is a node and edges pass one node's result into the next. That structure is what makes product fidelity tractable: the pack shot, the specifications, the logo and the style are four things the graph holds separately, so you change one without disturbing the other three and re-run everything downstream.

You brief an agent in plain language and it wires the graph and runs the models. Every step stays open, so the prompt behind a nearly-right shot is right there to correct. You direct; the agent produces.

The models are on one subscription. Melius spans a library of leading models for image, video and audio, and agents route each step to the strongest one for that step 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.

For related workflows, see building a static ad from a style reference and choosing a Nano Banana model.

Frequently asked questions

Have a pack shot and a product page? Open Melius and build the canvas on one SKU first.