Iteration, variations and prompt logs

Video lecture

Iteration, variations and prompt logs

10 chapters · about 8 min · full transcript

Coming soon

Chapter 1 of 10

Iteration and prompt logs

Great image, can't repeat it?
The iteration loop
Diagnose before you change
A log that takes seconds

The narrated lecture is in production

Every chapter is scripted and ready. Browse the chapters and read the full transcript now — the video will appear here when it’s published.

Chapters

Transcript of the narration, chapter by chapter.

0:00 Iteration and prompt logs

Have you ever made a great AI image, and then a week later, when the client asks for three more just like it, you can't get anywhere close? You've lost the prompt, you're not sure which setting you used, and you're back to pulling the slot machine lever. In this lecture, you'll learn the iteration loop that professionals use, how to diagnose before you change anything, how to use seeds, variations and conversational edits, and how to keep a prompt log that takes seconds to fill in. By the end, you'll be able to reproduce your best work, explain your process to a client, and stop wasting credits.

0:47 Why a process

Why does iteration need a process? Because image models are probabilistic. Each run is a sample. If you change several things at once, you can't tell which change helped. And if you don't write things down, you can't repeat success. Think about a chef developing a recipe. They don't throw in five new spices and hope. They change one ingredient, taste, and write it down. That's how a lucky dish becomes a menu item. Your prompts work the same way. There's also a business reason. A log is evidence of your creative process, which helps with client trust, disclosure questions, and even copyright discussions, where your human creative decisions matter.

1:35 The iteration loop

Here's the loop in ten steps. Brief: what is the image for? Draft: a seven-part prompt plus your style snippet. Explore: generate a batch of four to eight. Select the one or two closest to the brief. Diagnose exactly what's wrong. Adjust one main variable. Refine with variations, a fixed seed or references. Edit with inpainting, expanding or upscaling. Run quality checks for detail, rights, bias and brand fit. And log it. The key moment is diagnose. Wrong composition? Add shot type and placement. Wrong mood? Change lighting and palette, not just mood words. Unwanted text? Add a constraint or inpaint it away. Match the fix to the failure.

2:22 Seeds, variations and edits

Let's talk about the controls. Many tools expose a seed, which is the starting noise. Keep the seed fixed and change one word, and you see that word's effect clearly. Other tools offer subtle and strong variations, or remix. Use strong variations while exploring and subtle ones while refining. Assistant-built models let you edit by instruction. Make the lighting warmer. Move the cup to the left. That's fast, but it can drift, because each edit may quietly change other details, like the face or the label. So check the whole image after every step, and keep your best versions. One caution. When a vendor updates a model, the same seed and prompt may give a different result, so your saved final files are the real record.

3:17 Worked example 1: four runs, one change each

Worked example one. A simple product scene for a café's Instagram. Version one, the baseline prompt. Composition is good, but the light feels cold. Version two changes only the light, from cool daylight to warm late-afternoon light from the left. The mood is fixed, but the cup handle is distorted. Version three keeps the seed and adds a product reference at medium strength. The handle is correct now, but the background got busier. Version four adds soft, low-detail background on the right third. Clean copy space. Selected. Then one edit pass: inpaint the spoon, expand the top for a nine by sixteen story. Four runs, each with one change. And the log tells us it was the reference, not the prompt, that fixed the handle.

4:11 Worked example 2: a studio's shared log

Worked example two, a business scenario with illustrative numbers. A two-person studio in Abu Dhabi produces about forty social images a month for a hotel group. Before they had a log, repeat requests were painful. When the client asked for more images like the pool shot from March, the team spent hours trying to recreate it. So they started a shared log. One row per selected image, not per generation. Tool, model version, prompt, references, seed, edits, reviewer, and provenance notes. The next time the client asked for a match, they filtered the log, copied the prompt and references, and had a matching set in under half an hour. The log also answered the client's question about which images were AI-generated, instantly.

5:04 Watch me do it: a 60-second log entry

Watch me do it. I'll fill in a log entry for a finished image, and time it. I paste the asset ID and date. I pick the tool and model version from a dropdown. I paste the prompt I actually used, not the first draft. I list the reference file and its strength. I note the seed, if the tool shows one. Under edits, I write three short phrases: inpaint handle, expand top, upscale two times. I add my initials as reviewer. Under rights, I write no people, no logos. And under provenance, I note that Content Credentials are kept and whether a platform label is needed. Done. That took about sixty seconds, and it saves hours later.

5:55 Stop rules

Let's talk about managing time and cost, because iteration can become an endless slot machine. Set three rules. First, a time box. Twenty minutes of exploration per asset, then stop and re-read the brief. Often the problem is the brief, not the prompt. Second, a batch budget. A maximum number of generations per asset, and for paid work, agree it with your client. Third, the three strikes rule. If the same defect survives three targeted changes, stop prompting and switch mode. Use a reference, edit the region, composite a real photo, or try a different tool. These rules protect your margins and your sanity.

6:40 Common mistakes

Common mistakes. Changing five things at once and learning nothing. Logging every generation, which makes the log too big to use. Logging nothing, which makes success unrepeatable. Letting conversational edits drift without checking the whole image. Relying on seeds for long-term reproducibility when models get updated. And forgetting the context: which plan, which private mode, which reference files. If a client or a platform ever asks how an image was made, a vague memory won't help you. Your log will. Keep it simple enough that you'll actually fill it in, every time.

7:20 Recap and try this now

Let's recap. Iterate with a loop: brief, draft, explore, select, diagnose, adjust one variable, refine, edit, QA and log. Match each fix to the specific failure. Use seeds and subtle variations for controlled refinement, check the whole image after every conversational edit, and keep final files as the real record. Set a time box, a batch budget and the three strikes rule. Here's your try this now. Pick one real brief. Set a twenty minute timer. Run the full loop, changing one thing per run, and fill in a log row for your selected image using the template in the lesson. Then note which single change helped the most.

Professional results come from iteration

A single prompt rarely produces the final image. Professionals iterate deliberately: they change one thing at a time, keep what works, and record everything so results can be reproduced and explained to clients.

The iteration loop

1. Brief       → What is this image for? Format, placement, message.
2. Draft prompt → Seven-part structure + style spec.
3. Explore     → Generate a batch (e.g. 4–8 variations).
4. Select      → Pick 1–2 closest to the brief.
5. Diagnose    → What's wrong? Composition, lighting, detail, style?
6. Adjust      → Change ONE main variable.
7. Refine      → Variations, seed control, reference images.
8. Edit        → Inpaint, outpaint, retouch, upscale.
9. QA          → Details, rights, bias, brand fit.
10. Log        → Save prompt, settings, references, final.

Diagnose before you change

When an image misses, identify the specific failure:

Problem	Likely fix
Wrong composition	Add shot type, angle, placement; use a structure reference
Wrong style	Strengthen style spec or style reference
Wrong mood	Adjust lighting and palette, not just mood words
Extra/missing objects	Simplify the prompt; state counts; inpaint
Faces/hands distorted	Regenerate variations, adjust pose description, inpaint
Unwanted text/logos	Add constraint; remove via inpainting
Too generic	Add specific details, textures, imperfections

Seeds and variations

Many tools expose a seed — the starting noise pattern. Keeping the same seed while changing one prompt element lets you see that element's effect more cleanly. Other tools offer "vary subtle / vary strong" or "remix" options instead. Use:

Strong variations during exploration.
Subtle variations and fixed seeds during refinement.

Note that model updates can change results even with the same seed and prompt, so do not rely on seeds for long-term reproducibility — keep the final images and references.

Conversational editing

Assistant-style tools let you refine by instruction ("make the lighting warmer", "move the cup to the left"). This is efficient but can drift: each edit may subtly change other details. Check the whole image after each step and keep the best versions.

The prompt log

A prompt log turns trial and error into a reusable asset:

Prompt log entry
Project / asset:     Café launch — IG 4:5 — hero image
Date:                2026-03-14
Tool / model / mode: [tool] [model version] [private mode]
Prompt:              [full text]
Negative/constraints: [text]
References:          style_ref_01.png (strength medium)
Seed / settings:     [seed] [aspect ratio] [quality/steps]
Selected output:     cafe_hero_v3_04.png
Edits:               inpaint cup handle; outpaint top for text; upscale 2×
Rights/QA notes:     no people; no logos; reviewed by AB

Benefits:

Reproduce and extend a look for future posts.
Explain your process to clients (transparency builds trust).
Support any future questions about how an image was made — useful for disclosure, provenance and copyright discussions.
Train team members.

Managing time and cost

Iteration can become an endless slot machine. Set limits:

Time box: for example, 20 minutes of exploration before reassessing the brief.
Batch budget: a maximum number of generations per asset.
Stop rule: if the model cannot do it after structured attempts, switch approach (different tool, reference image, edit, photo, or illustration).

Worked example: a product scene

Brief: 1:1 ad image of a reusable water bottle on a rock by a mountain stream, morning light, space for a headline.

Batch 1: nice scenes but the bottle design is inaccurate (the model does not know the product).
Diagnosis: product accuracy cannot come from text alone.
Approach change: generate the scene without the bottle, then composite the real product photo in the design tool, matching lighting and adding a shadow. (If the product is shown in a scene it was not photographed in, ensure the ad does not imply claims about the product that aren't true.)
Final: scene upscaled, real product composited, headline added in the brand font.

Common mistakes

Changing many variables at once.
Not saving prompts and settings.
Endless rerolling without diagnosing.
Trying to force a model to render an exact real product from text alone.

Hands-on: a prompt log you can actually maintain

A log only works if it is fast to fill in. Use a shared spreadsheet or database with these columns, one row per selected output (not every generation):

asset_id | date | brief_link | tool | model/version | mode (private?) |
prompt | constraints | references (+strength) | seed/settings |
selected_file | edits (inpaint/expand/upscale) | reviewer | rights notes |
provenance (Content Credentials kept? label needed?) | status

Or, if your team works in a repository or automation tool, store each entry as JSON next to the final file:

{
  "asset_id": "cafe-launch-ig45-hero",
  "tool": "[tool name]",
  "model": "[model/version]",
  "prompt": "Editorial photograph of ...",
  "references": [{"file": "style_ref_01.png", "strength": "medium"}],
  "edits": ["inpaint cup handle", "expand top 20% for headline", "upscale 2x"],
  "reviewer": "AB",
  "provenance": {"content_credentials": true, "platform_label": "not required: stylized"},
  "status": "approved"
}

Worked example: one change at a time (before/after log)

Run	Single change	Result	Keep?
v1	Baseline seven-part prompt	Good composition, light too cold	Partly
v2	"cool daylight" → "warm late-afternoon light from the left"	Mood fixed, cup handle distorted	Yes
v3	Fixed seed, added product reference at medium strength	Handle correct, background busier	Yes
v4	Added "soft, low-detail background on the right third"	Clean copy space	Selected
Edit	Inpaint spoon, expand top for 9:16 story	Final	Logged

Four runs and one edit pass, each with a single change, took about 15 minutes (illustrative). Without the log, the team would have forgotten that the reference, not the prompt, fixed the handle.

Stop rules

Three strikes: if the same defect survives three targeted changes, switch mode (reference, edit or composite) or switch tool.
Time box: 20 minutes of exploration per asset, then re-read the brief.
Budget: a maximum number of generations per asset, agreed with the client for paid work.

Summary

Iterate through a structured loop, diagnose before changing, use seeds and variations deliberately, keep a prompt log, set time and batch limits, and switch approach when prompting alone cannot deliver.

Iterate with a loop — brief, draft, explore, select, diagnose, adjust one variable, refine, edit, QA, log.
Diagnose the specific failure and match the fix (composition, lighting, reference, edit) rather than rewriting everything.
Fixed seeds and subtle variations support controlled refinement; model updates can break seed reproducibility.
Log one row per selected image, including references, edits, reviewer and provenance decisions.
Use stop rules: a time box, a generation budget and three strikes before switching mode or tool.

Pick one asset brief and run the full iteration loop with a strict 20-minute time box, recording each step in a prompt log entry. Note which single change improved the image most.

Enrol for free to save your progress

Reading is always free. Enrol to keep your place, take the final assessment and earn a verifiable certificate.

Iteration, variations and prompt logs