Pixolm

Pixolm brand concept visual.

The first AI image is rarely the finished asset. The next request is usually much more specific: change the setting, soften the light, or leave room for a headline.

I'm building Pixolm full-time as a solo founder, and this is the workflow I wanted to test: can a conversation carry an image from a basic starting point toward a useful marketing composition?

3D product concept visual, not a live interface screenshot.

Here are two real experiments from Pixolm Agent. Each uses three separate generation requests, with the previous result as a reference for the next step.

1. Skincare: a clean product shot becomes a campaign concept

Step 1 — Establish the product. A fictional mint-green serum bottle on a white background. No brand, label, or text.

Step 2 — Give it a setting. Using the first image as a reference, I asked for shallow water, transparent glass blocks, and a calm mint palette. The direction becomes much clearer when you can react to an existing image.

Step 3 — Make space for the message. I asked to move the bottle toward the lower right and leave the upper left open for copy. The result is a layout concept; the headline still needs to be added.

2. Coffee: a menu image gets a seasonal direction

Step 1 — Start simple. An unbranded iced latte against a clean background.

Step 2 — Introduce a mood. I used that result as a reference and requested a wooden table, warm sunlight, a few chestnuts, and cream linen for an autumn cafe scene.

Step 3 — Refine for a social post. The follow-up asked for fewer props, a smaller drink in the lower right, and more open space above it for a seasonal menu announcement.

What this taught me about the workflow

The useful part was being able to make one concrete decision at a time: first the subject, then the setting, then the composition. I used Korean instructions for these examples; the descriptions above summarize them in English.

There is still a real limitation: reference images help guide the next result, but they do not lock every detail. Bottle proportions, glass shape, and surface details can change. These are fictional product concepts, not proof of exact SKU preservation or finished client deliverables.

All six images were generated in Pixolm Agent using GPT Image 2.5 Sunburst, 1K, High, at 4:5. They are actual outputs, not interface mockups.

I'm building the agent alongside manual creation and targeted editing because different stages need different amounts of control.

The logo, workspace, and gallery visuals were created separately with Codex ImageGen for product presentation. The six skincare and coffee images above are actual Pixolm Agent outputs.

If you make visuals for a small brand or a client, which matters most after the first image: keeping product details, changing the scene, or getting the layout right? I'd love to hear where your workflow gets stuck.

Try Pixolm: https://pixolm.com/

More examples and updates: https://x.com/Pixolm · https://www.youtube.com/@PixolmAI