# The Art Direction Bake-Off Kit

See real options on your own content before you commit to a look. One afternoon, a few dollars.

Model notes are as of October 2026. Models change fast, so rerun the test before you standardize.

## 1. Before you generate

- [ ] **Write a one-page brief.** What the images are for, who sees them, and five to ten hard rules (colors allowed, colors banned, no text in images, where empty space goes).
- [ ] **Pick three real graphics** you actually need. Test on your content, never on sample prompts.
- [ ] **Define three to five styles** in one sentence each. Make them different enough to argue about.
- [ ] **Choose three or four image models.** One account that reaches several models keeps costs and logs in one place.
- [ ] **Set a budget.** Ours was $5.64 for 78 images.

## 2. While you generate

- [ ] **Same prompt, every model.** Paste it exactly. No extra words.
- [ ] **First outputs only.** No rerolling until the scoring is done.
- [ ] **Log every image:** prompt, model, cost. That log is your receipt.
- [ ] **One contact sheet per style,** every model side by side.
- [ ] **Run long batches in the foreground.** Background jobs can die when the session ends.

## 3. Score it

Score each image 1 to 5, multiply by the weight. Total possible: 60.

| Criterion | Weight | What a 5 looks like |
|---|---|---|
| Stops the scroll | 3 | You would stop on it in a feed |
| Idea | 3 | One clear idea, not a literal picture |
| Line and craft | 2 | Confident, consistent, looks made on purpose |
| Rule following | 2 | Every rule in your brief, kept |
| Composition | 1 | Room where your headline or text goes |
| Fits the family | 1 | Sits next to the others without clashing |

| Style | Model | Graphic | Scroll 3x | Idea 3x | Craft 2x | Rules 2x | Comp 1x | Family 1x | Total | Rule breaks |
|---|---|---|---|---|---|---|---|---|---|---|
| | | | | | | | | | | |

## 4. Decide

- **Reliability beats one lucky image.** A model that is great once and wrong twice loses.
- **Count rule breaks** (stray words, banned colors, a doubled character). They tell you what each model ignores.
- **Favor styles that look good on every model.** They do not depend on one vendor staying good, cheap or available.
- **Check you can repeat it:** cost per image, speed, and whether the model is reachable the same way next month.

## 5. Model habits we saw (October 2026)

| Model | What we saw |
|---|---|
| Gemini 3 Pro | Best craftsmanship and color fidelity |
| GPT Image 2 | Most consistent, best visual ideas, no rule breaks we saw |
| Grok Imagine Quality | About half Gemini's price per image; strong in several styles; weaker at following "no" rules |
| Recraft V4.1 Vector | The only one that returns real SVG. Hand it a reference and it redraws the reference, so use it without one, for the final vector pass |

## 6. The question to ask first

If the output keeps landing below your bar, ask whether you are using the wrong tool before you rewrite the prompt.

From Emmanuel Quintero's Lab Notes: "78 Images, 4 Models, $5.64: How I Chose My Art Direction."