How the audition works

Approach A · the five neighbors of the first picture

Approach B · the same five, repainted with the reference

Start with words

Everything starts with a short written brief. It describes the picture we want: a lighthouse on rocks. Its beam cuts through TV static. The static melts into clear sky. Two lines of text sit on the picture. There is no lighthouse image yet, only words. Then we add one reference image. It can be anything: a painting, an old photo, a glitchy portrait. The question we are asking is simple. What does this look do to our picture? And what other looks live right next door to it?

Step 1. Claude writes six recipes

Claude is the AI writer in this system. It reads the brief and looks at the reference image. It also gets a short brand sheet that fixes the fonts and colors of the text. Then it writes six recipes. Each recipe is a structured list: the mood, the medium (paint, photo, print), the colors as exact hex codes, where each thing sits in the frame, and the text. Recipe one is the canonical: the brief taken literally. Here is the surprise. Claude describes the reference's medium and colors inside that first recipe. So the first picture already carries the reference's look. It is never a plain lighthouse waiting to be styled. The other five recipes are neighbors. Same lighthouse, same layout, but one or two things change: the medium, the colors, the light, the era, the texture, or how abstract it is. That is why the six pictures look so different. They fan out from a look, not from a blank.

Step 2. Ideogram 4 paints them

Ideogram 4 is the image model that turns recipes into pictures. It runs in ComfyUI on the Spark, a small computer with a big graphics chip. Ideogram never sees the reference image. It only reads Claude's recipe, which describes the reference's medium and colors in words. All six pictures use the same random seed, so the layout stays steady and only the look changes. Each one takes about 25 seconds.

Step 3. Claude writes a repainting note

Now Claude writes one more thing: a note for a different tool, an image editor. The note says how to repaint a picture so it looks like the reference. It names only the surface: the paint, the paper, the finish, the colors, the light. It must never name what the reference is a picture of. If it did, the editor would paste that thing into our lighthouse scene. The code adds one rule at the end: keep the scene from picture one.

Step 4. No text yet, on purpose

Here is a lesson we learned the hard way. Image models treat letters like any other shape. They tint them, smear them, or invent new ones. Our first version let Ideogram paint the text and then tried to hide it from the editor with a blur and put it back afterwards. It left ghosts and doubled words. So now the pictures are made with no text at all. Claude's recipe still keeps a quiet, empty corner where the text will go, described as plain background, so the picture leaves room for it.

Step 5. Qwen repaints

Qwen Image Edit is the editor. It also runs on the Spark. It takes two pictures: our clean, text-free lighthouse and the reference. It repaints every surface of the lighthouse picture in the reference's material, following Claude's note. The layout stays. The skin changes. If the reference is mostly a person or a face, Claude asks for palette mode instead. Then the editor only sees a blurred wash of the reference's colors, so no face can sneak in.

Step 6. Set the text last, then sign it

Back on the Mac, a small program sets the two lines of text onto every finished picture, both the Ideogram ones and the repainted ones, using the real brand fonts: Jost Bold for the headline, Figtree for the smaller line. It measures the picture under the text and picks Warm White or Charcoal so the words stay readable. If the ground is busy, it lays a soft shadow under the text first. Every picture in a run gets identical type in the same place. Then the K mark goes in the bottom corner. Every picture also keeps its full ComfyUI recipe inside the file, so you can drag it back into ComfyUI and rebuild it.

The end result

For every reference image we get two answers. Approach A is the neighborhood: six looks written in words and painted from those words. Approach B is the reference's real surface, moved onto those six pictures. Six pairs per reference. One hundred thirty references. One page per run, one contents page per ten, and a table where you write down which one wins. The brief never changed. Only the look did.

We run Ideogram 4 ourselves

ComfyUI's built-in Ideogram node sends plain text and lets Ideogram rewrite it. Running the open model ourselves lets us send the exact recipe. The open model is for non-commercial use, so this is for choosing a look, not for final client art.

The recipe is the plan

Exact colors, a medium, the light, and where each thing sits. Change one line and you get a true neighbor. Words are open to interpretation; color codes are not.

Two approaches, on purpose

Approach A changes medium and colors in words. Approach B paints over them with the real reference. They disagree by design. A shows the neighborhood. B shows what the reference's surface really does.

One seed for all six

A seed is the random starting point. Keeping it fixed keeps the layout steady, so the only thing that changes is the look.

Text is set last, by a program, not a model

Image models paint letters as shapes and get them wrong. So no picture is made with text. The brand text is typeset at the end with the real font files, identical on every picture in a run.

The brand owns the text

Jost Bold for headlines, Figtree for the smaller line, Side Car colors. The reference can change the size and placement of text, never the font or color.

Why does the first picture already look like the reference?

Because Claude looks at the reference before writing recipe one, and describes its medium and colors inside that recipe. Ideogram cannot see images. The reference reaches it only as words, and those words are in the first recipe from the start. The five neighbors then change things from that already-styled picture. That is where the variety comes from.

Then what does Approach B add?

The real surface. Approach A carries the reference as a description, like matte paper and a color list. Approach B repaints with the reference's actual texture, edges and color mix. Look at the static in the demo: same layout, different skin.

Why did some repaints paste the reference's subject in?

In early runs the editor was shown the reference itself, even when the reference was mostly a person. Now Claude switches to palette mode for those, a blurred wash of colors, and the problem stops. Old runs can be redone one at a time in about six minutes.

Is the text ever painted by a model?

Not any more. Early runs let Ideogram paint it and then blurred it out for the editor; that left ghosts and doubled words. Now every picture is made without text and the brand text is set on top at the end, so the letters are the same pixels on all twelve pictures of a run.

Can I rebuild any picture?

Yes. Every PNG file carries its full ComfyUI recipe. Drag it onto the ComfyUI canvas on the Spark. Repainted pictures from before 4 September 14:15 need their two input images uploaded again from the raw folder; later ones already have them.

What did it cost?

About eleven minutes and one dollar of Claude per reference. One hundred thirty references took about a day, including two restarts that picked up where they left off.

• The Ideogram 4 open model is non-commercial. Final client art needs a licensed re-render.
• Ideogram cannot see images. The reference reaches Approach A only through Claude's words.
• Approach B pulls all six pictures toward one material. That is by design; it is not a second fan-out.
• The rig ranks nothing. The verdict is yours.
• The report pages are served from the MacBook. The Mac has to stay awake.