# Composition checklist and comparison worksheet

A blank worksheet for planning and reviewing your own image project. It contains no experimental results. The A/B exercise proposed in the article has not been performed, and the article's single scene illustration is not part of it.

Use **pass / fail / uncertain** for each applicable check. Give an observation, not just a rating. Leave unreviewed items blank, and keep uncertainty separate from a pass. If a check does not apply, explain why and revise the checklist before reviewing outputs.

## 1. Define the brief before generating

- Project and intended use:
- Main message:
- Required objects, quantities, and attributes:
- Required spatial relationships, using image-left and image-right:
- Required text and where it must fit:
- Target dimensions, display context, and permitted crop:
- Requirements that cannot be traded away:
- Negotiable style preferences:
- What a reviewer should be able to recognize:

Write the actual headline, date, and location here. Mark fictional practice copy clearly.

```text

```

## 2. Plan a narrow prompt comparison, if needed

Skip this section for a single-image review. Complete it before generating a comparison.

- Question being tested:
- Exact sentence or constraint being changed:
- Planned outputs for each version:
- Generation order:
- Review criteria agreed before seeing results:
- How all outputs and failed attempts will be retained:
- How reviewers will be kept unaware of version labels, if applicable:

For the article's proposed exercise, **B is the complete practice template in Section 4**. **A removes only this sentence**:

> Keep the lampshade's material recognizable, and avoid turning it into a large patch of overexposed white.

Copy and save both complete prompts before starting. Keep every other sentence unchanged. This tests that one added constraint, not long prompts against short prompts in general. If you choose a different change, record it above instead.

Four outputs per version is an optional teaching-budget example from the article, not a required or statistically validated sample size. Enter your actual plan, retain all results, and keep conclusions limited to that plan.

### Inputs to keep the same

Record only settings actually exposed by the tool. Use “unavailable” for a model version or seed that cannot be verified. Do not guess hidden settings.

| Input | Recorded value | Same for A and B? / reason for any difference |
|---|---|---|
| Tool and date | | |
| Visible model name | | |
| Verifiable model version, or unavailable | | |
| Requested dimensions or aspect ratio | | |
| Other visible generation settings | | |
| Reference images and their attachment order | | |
| Seed plan, or unavailable | | |
| Complete prompt A, saved file or pasted text | | |
| Complete prompt B, saved file or pasted text | | |

When seeds are available, paired seeds can be used within the same model version and settings. Matching seeds do not establish comparability across different models. Fixed inputs do not guarantee that other image details stay unchanged.

## 3. Keep every attempt and every output

Add rows as needed. Preserve originals, including images that fail the brief. Record a request that returned no image separately from an image that was generated but contains errors. Do not silently replace failed attempts with reruns or omit them from the record.

| Attempt label | Version | Seed or unavailable | Returned output label(s), or no image | Actual dimensions | Request failure or other observation | Retained file location |
|---|---|---|---|---|---|---|
| | | | | | | |
| | | | | | | |

- Deviations from the plan, including extra attempts and their reasons:
- Original files preserved at:
- Diagnostic copies, crops, or edits stored separately at:

## 4. Review each output

Duplicate this section for **every** returned image. Preserve the full original alongside any diagnostic copy. Do not judge an entire version from its best image.

- Output label:
- Reviewer:
- Review date:
- Version hidden during review?:
- Original dimensions:
- Actual display size and viewing context checked:
- Actual text and crop tested:

| View / criterion | Pass, fail, or uncertain | Visible evidence and reason | Next action |
|---|---|---|---|
| Original: required objects and quantities are correct | | | |
| Original: attributes and spatial relationships match the brief | | | |
| Original: important structures and contours are recognizable | | | |
| Original: no required part is cut off | | | |
| Original: lampshade highlights retain recognizable material, if required | | | |
| Intended display size: the reading subject is recognizable | | | |
| Intended display size: the emphasis serves the brief | | | |
| Grayscale diagnostic: supporting light areas do not obscure the intended hierarchy | | | |
| Return to color: light–dark and color relationships meet this project's requirements | | | |
| Complete layout: the actual headline and event details are legible | | | |
| Complete layout: the final crop preserves required information | | | |

A grayscale copy is a diagnostic view, not a pass/fail rule for every image. Color can carry important distinctions. Grayscale, blur, and manually marked hotspots are not measured eye-tracking heatmaps.

### Optional open-ended reviewer notes

Ask without suggesting the expected answer:

- “What is this image mainly about?” Response:
- “Which areas draw your attention?” Response:
- Other uncertainty or disagreement:

These are subjective reports, not measured first fixations. They do not establish click-through or conversion rates.

## 5. Record a revision and check for regressions

Keep the pre-edit version. A request to preserve an area is not proof it remained unchanged.

- Source output label:
- Specific error to address:
- Exact edit instruction:
- Intended preserved features:
- Result label and retained original:

| Check | Pass, fail, or uncertain | Observation |
|---|---|---|
| The target error improved | | |
| Previously correct content remains correct | | |
| Previously accepted positions and text space remain usable | | |
| Color and light remain suitable for the brief | | |
| The complete layout still meets delivery requirements | | |

- Accept, revise further, return to the previous version, or change method:
- Reason:

## 6. Write a conclusion only after review

- Planned attempts / completed attempts:
- Returned images / requests that returned no image:
- Confirmation that every returned image was retained and reviewed, or a description of the gap:
- Observed differences between A and B, tied to output labels and criteria:
- Failures, uncertain judgments, and reviewer disagreements:
- Uncontrolled differences or deviations from the plan:
- What this exercise supports for this project:
- What it cannot establish:
- Next change or follow-up check:

Do not turn a small exercise into a claim of statistical significance, general model superiority, or a universal prompting rule. Recheck when the model, subject, settings, or intended use changes. If you summarize counts, calculate them from the completed records and report uncertain judgments separately.

## 7. Delivery decision

- Required content checks completed:
- Actual display size and crop checked:
- Actual headline, date, and location checked:
- Remaining failed or uncertain requirements:
- Decision and responsible reviewer:

An attractive image is not automatically ready to deliver. Keep an unmet requirement visible until it is corrected or the project owner explicitly changes the brief.