Flat wrap design with AI: 4 pipelines tested side-by-side

In "5 AI workflows for first-meeting wrap concepts" and "3 ways to generate AI vehicle wraps using a custom-trained model" we tested workflows that produce 3D mockups — the wrap rendered onto a photo of the vehicle, the thing you'd email a client to say "this is what your van would look like."

This post tackles a different job: flat panel design — the print-ready 2D artwork file a wrap shop physically cuts and applies. No vehicle in the image, no 3D, no shadows. Just the wrap design itself, sized to match a real vehicle panel's proportions, ready to send to a vinyl printer.

That sounds easier than a 3D mockup. It isn't.

---

Flat panel vs vehicle mockup: different job, different pipeline

A 3D mockup is forgiving. The wrap can be partially obscured by the angle. Text can sit on a body panel that curves away. The viewer's job is to imagine the finished vehicle, and the AI's job is to provide a believable-enough preview.

A flat panel design has nowhere to hide. Every text character has to be perfectly readable. The design has to flow across the actual physical proportions of a vehicle panel (a van's driver side is roughly 3:1 to 4:1 — long and narrow). No curves to absorb mistakes. And critically, the image must be a clean 2D graphic — not a photo of a sign or billboard, not a card with a white border, not a poster pinned to a wall. Just artwork that fills the canvas edge-to-edge, ready for vinyl cutting.

The problem: image generation models like Flux are trained mostly on photographs, posters, and cards. When you ask for a "flat vehicle wrap panel" they default to producing one of those familiar things — usually with a frame, a shadow, or worst of all, an actual vehicle in the background. Getting them to output a true print-ready flat panel takes deliberate engineering.

We built four pipelines that try to solve this, with increasing levels of complexity, and ran them all against the same fictional brief.

---

The test: same fictional brand, flat-design only

We reused the brand from our previous posts — Pampered Paws Mobile Grooming — so the inputs are directly comparable.

Heads up: "Pampered Paws Mobile Grooming" is entirely fictional. The brand name, logo, phone number, website, email, and tagline were all fabricated or AI-generated for testing. No real business was involved.

• Brand: Pampered Paws — mobile pet grooming
• Tagline: "We Come to You!"
• Colors: teal, coral pink, white
• Finish: gloss
• Contact: phone, website, email

Every pipeline got exactly the same inputs. Each pipeline was run 4 times to test consistency — the runs grids below show every output, no cherry-picking.

---

Workflow 1: production baseline — render once, edit once

The idea. This is the current production flat pipeline. Concept LLM plans the layout, Flux renders the panel in a single shot, then an editor LLM reads the rendered image and applies one round of corrections. That's it — no verification of the final output, no retry, no fallback.

What happened. Fastest and cheapest. But Flux only gets one chance to render the text correctly, and the single edit pass doesn't catch much. Brand names came out close but rarely perfect. Phone numbers were garbled in most runs. Website domains drifted to similar-but-wrong strings (pamperedpaws.com instead of pamperedpawsmobile.com). And the output was always rendered as a near-square business-card shape with a soft white border around the design — not the long horizontal panel a real van side actually needs.

The cheapest pipeline, but also the lowest quality — usable for a quick sketch, not for a vinyl shop.

---

Workflow 2: add a verify-fix loop for text

The idea. Same first pass as Workflow 1, but instead of stopping after one edit, add a verify-fix loop: a vision LLM reads every text element on the panel character-by-character, compares to the brief, and asks the editor to fix any mistakes. Loop until text is exact, or until 3 iterations.

What happened. Text accuracy jumped dramatically. The verify-fix loop reliably caught misspelled brand names, garbled phone numbers, and wrong websites. But each correction pass also slightly degraded the design: backgrounds got busier, color palettes drifted, sometimes a mascot character was hallucinated where the brief never asked for one. And the output was still near-square — the editor's underlying API only supports a 1.5:1 aspect ratio, so the wide horizontal layout was always compressed back into a card shape.

A clear improvement over the baseline on text fidelity, at meaningfully higher cost and longer run time — and the wrong aspect ratio for a real wrap panel.

---

Workflow 3: graphics-first, then add text separately

The idea. What if the problem is that Flux can't do graphics AND text well at the same time? Try splitting them: in the first pass, prompt Flux for shapes and gradients only — no text, no logo, just a clean "reserved zone" where text will go later. In a second pass, ask Gemini to add the text into that reserved zone. Then add the logo. Then run the verify-fix loop from Workflow 2.

What happened. Mixed results. When it worked, the graphics were cleaner than Workflow 2 (because Flux wasn't stretching itself to render text on top). But the text-add step was unreliable — sometimes Gemini placed text correctly, other times it ignored the reserved zone and crammed text into a corner, and occasionally it just rendered the wrong character set. The worst case: in some runs, Flux ignored the "no text, no vehicle" instruction completely and produced a full 3D vehicle mockup with the wrap already applied.

About the same quality as Workflow 2 on a good day, but with much higher variance run-to-run — the kind of pipeline that produces both your best result of the day and your worst.

---

Workflow 4: vehicle-aware, 21:9 aspect, Gemini-only edits — the FINAL

The idea. Three big changes from the previous workflows:

1. Vehicle geometry analysis up front. Before any generation, a vision LLM reads the vehicle photo and identifies door positions, wheel-well locations, body lines, and recommended text zones. This information is used internally to plan the design but is never included in the Flux prompt — keeping vehicle words out of the prompt is the key to not getting a vehicle rendered.

2. Real panel proportions. Render at 1440×617 pixels (21:9 ratio) — the actual proportions of a van's driver side. No more near-square cards.

3. Abstract zone language + Gemini-only editing. The concept LLM is forbidden from using the words vehicle, car, van, truck, panel, door, wheel, bumper, poster, sign, billboard, card. It describes the design as "left zone / center zone / right zone" — pure abstract graphic-design language. And editing uses Gemini only, never OpenAI: OpenAI's image edit API only supports 1:1, 3:2, and 2:3 sizes, so it would reshape our 21:9 wide panel back into a 1.5:1 box. Gemini natively supports 21:9.

What happened. Everything fell into place. The output is a clean, edge-to-edge flat 2D graphic at real vehicle-side proportions. Text comes out correct on almost every run. The design quality is the highest of any pipeline. And surprisingly, this was also one of the cheapest pipelines and the fastest — because Gemini-only editing skips the OpenAI fallback chain and the verify-fix loop usually converges in fewer iterations on this aspect ratio.

This is what we're rolling out to production as the new flat panel pipeline. The old pipeline stays available so we can track which workflow produced each design.

---

Side-by-side: the best output from each pipeline

Same fictional brief, four pipelines, one sample from each. The shift in panel shape from Workflow 1 → Workflow 4 is the single most visible change.

---

What we learned

1. Words matter more than rules.

The biggest single quality jump came from removing the words "vehicle", "car", "van", "truck", "panel", "door", "wheel" from the concept-LLM prompt. Even when those words appeared only in the instructions (not the final Flux prompt), Flux picked them up and rendered the wrong thing. Reframing the design as "left zone / center zone / right zone" got Flux to produce flat graphics consistently. Negative rules alone ("do NOT render a vehicle") aren't enough — Flux still slips up. Removing the trigger words entirely is what works.

2. Match the rendered aspect ratio to the editor's supported sizes.

Wide aspect ratios get destroyed by the wrong editor. OpenAI's image edit API only supports 1:1, 3:2, and 2:3 — so a 21:9 wide panel gets reshaped to 1.5:1 and the layout collapses. Gemini supports 21:9 natively. The right pipeline decision isn't "which editor is better in the abstract" — it's "which editor preserves the dimensions you actually need."

3. Verify-fix loops fix text, but only if they don't redo the design.

Adding a verify-fix loop (Workflow 2) lifted text accuracy substantially, but the same loop sometimes degraded the visual design. The fix was to make the fix-prompts as narrow as possible ("change dosn't to doesn't, keep everything else identical") instead of letting the editor decide what to change. Targeted edits preserve the original design; open-ended ones rewrite it.

4. Cheaper and better can coexist.

Workflow 4 is both higher quality and cheaper than Workflows 2 and 3. The savings come from skipping OpenAI entirely (Gemini-only) and from the verify-fix loop converging faster when the underlying render is already good. The lesson: when you're tempted to add fallbacks "just in case," check whether the fallback is actively hurting you in the success path.

---

Where we're still improving

• Tagline rendering is still the weakest link — short text floats around the panel and the editor sometimes drops an apostrophe or a period. Considering a dedicated tagline-layout pass.
• Logo placement is consistent on the side panels but slightly cramped on the square rear panel. Tightening the per-panel layout prompts.
• Group-flat (driver + passenger + rear + front) generation works but uses one "hero" panel as a reference for the others to keep colors consistent. Run time scales linearly with panel count and we're looking at parallelizing the per-panel loop.

If you want to run your own brief through this pipeline, start a flat design in the app. The new vehicle-aware pipeline (Workflow 4) is the default for new flat designs going forward; the older one-render pipeline stays available, and every generated design is tagged with the workflow version that produced it so we can track quality over time.