An AI art iteration log: change one variable at a time
An AI art log is useful because it separates the variable being tested from everything else. Change one element at a time, compare results directly, and record conditions that the tool may or may not expose. Seeds, references, crops, and resolution should each be treated as separate experiments rather than blended into one impression of better or worse. This keeps credit where it belongs and prevents a lucky output from becoming a recipe that cannot be reproduced.
Set a baseline before guessing
Start with one approved baseline image or prompt and freeze everything except the planned variable. Record the prompt, aspect ratio, reference use if any, crop, export size, and any tool-reported seed or parameter. Use three or five candidates so one outlier does not dominate. Name the test, date, and criteria before viewing results. Criteria can include subject clarity, spatial balance, artifact visibility, and suitability for the intended crop. Without a baseline, changing composition and reference together makes it impossible to know which change helped.
Test composition separately
Composition changes include camera angle, subject scale, horizon placement, symmetry, negative space, and focal order. Change only one such element while keeping style, subject, reference, and format stable. Ask whether the new balance improves the intended use, not whether it looks more dramatic in isolation. If a stronger composition requires a wider frame, note that crop implications separately. A composition test should produce a decision: keep, reject, or modify the baseline.
Handle references carefully
A reference image can influence pose, palette, lighting, texture, or subject, but its effect is not always isolated or predictable. Keep reference version, placement, and any strength setting constant across the composition test. Compare reference-on and reference-off only as its own experiment. Record whether the output appears closer to the goal and whether it introduced unwanted elements. Do not assume the tool interprets the same reference consistently at different resolutions.
Seeds if available
Treat a seed as an observation, not a guarantee. If the tool reports or allows one, record it and test whether changing seed while holding everything else constant alters composition in useful ways. If no seed is exposed, label the field unavailable instead of implying hidden control. Some systems map seeds differently or ignore them after model updates. Useful experiments compare same-seed candidates against different-seed candidates without changing prompt or placement. Never promise reproducibility from a seed alone.
Crop as its own review
Cropping changes perceived composition, subject weight, and usability without regenerating the image. Test original, centered crop, and subject-first crop at the target placement. Compare whether important details survive, whether empty space feels intentional, and whether the crop hides flaws. Crop review should happen after composition selection, because a good full frame can fail in a required frame. Record dimensions and intended placement for each candidate.
Resolution and export checks
Run resolution and export as a separate stage after selecting image and crop. Record generation size, export pixel dimensions, compression, and any profile reported by the tool. Look for upscale artifacts, edge halos, text errors, color shifts, and loss of small detail. Compare original and exported images side by side because viewers usually judge the exported file. If the platform offers enhancement, test it separately and label it enhancement, not composition.
Comparison sheet
Export a comparison sheet with rows for variable, value, candidate label, and evaluator notes, plus columns for each result. Include scores for clarity, composition, artifact level, and crop fit, then a final selection and reason. Keep notes factual and describe what changed, not taste alone. Save the sheet with the baseline and exported images so a later test can repeat only the promising variable. This turns a folder of similar files into evidence for the next creative decision.
Checklist
- Freeze all settings except one documented variable.
- Record seed as available, unavailable, or tool-specific observation.
- Test crop separately after choosing a composition.
- Export filenames with variable and candidate labels.
- Keep a comparison sheet with criteria and final rationale.
Takeaway
An iteration log becomes reliable when composition, reference, seed behavior, crop, and resolution are tested separately. Small controlled changes reveal more than a stack of impressive but unexplained outputs.