Four Commands to Fix a Thumbnail: Draw More, Not Different
One line fixes most AI cover images: “Give me four versions in a single grid. Change only the background colour. Number them 1 to 4.”
One prompt, four options, pick the winner. That is the whole trick, and it works because it turns an impossible task – describe the perfect image in words – into an easy one: look at four things and choose.
Behind it sits a rule worth more than any single prompt: draw more, not different. Never start from a blank prompt. Keep the layout that already works and change what is inside it. Make one change per command, not five. And when something is wrong, do another pass rather than writing a longer prompt.
Four passes, not one long prompt
The four commands are short on purpose. They run in order, and each one leaves the previous result intact.
| Step 0 · Reference → | Pass 1 · Size → | Pass 2 · Edits → | Pass 3 · Subject → | Pass 4 · Grid |
|---|---|---|---|---|
| Start from something that works A thumbnail already winning in your niche, found in search |
Get the canvas right 16:9, 1280×720, full bleed, keep the eye path |
One object at a time Remove this, add that, scale the subject, move it |
Make it yours Your face, same lighting and angle, background untouched |
Choose, do not guess Four numbered variants, one thing changed between them |
| Structure only | Layout kept | One change per pass | Your photo | Pick a number |
Long prompts fail for a simple reason: when eleven instructions go in at once and the result is wrong, you cannot tell which instruction caused it. Four passes give you four checkpoints, and the fix is always obvious because only one thing changed.
The four commands, word for word
Here they are with the brackets you fill in, and – just as important – what each one is told not to touch.
| Command | The line | What it changes | What it must not touch |
|---|---|---|---|
| 1 · Size | “Redraw this reference at [16:9], 1280×720. Keep the layout and the eye path. Full bleed, no border.” | The ratio and the pixel size | Anything about content |
| 2 · Edits | “Remove [object], add [object], scale the subject to [60%] of the frame and move it [left].” | What is in the frame and how big | The lighting and the layout |
| 3 · Subject | “Replace the face with [my photo]. Keep the same lighting, same angle, same expression. Don’t touch the background.” | Who is in it | Everything around them |
| 4 · Grid | “Give me [4] variants in one grid. Change only the [background colour]. Number them 1 to 4.” | One variable, four ways | Everything else, deliberately |
Look at the fourth column. Every command contains an exclusion, and that is not decoration: “don’t touch the background” is what stops a face swap from redrawing the whole image. The same structure makes ordinary prompts reliable, as we set out in image prompts in five slots, and the bracket you fill decides the rest, as in the one part that decides whether a prompt works.
Command four is the one to memorise. Four variants in one grid costs a single generation, gives you a side-by-side comparison your eye can judge instantly, and ends with you naming a number instead of writing another paragraph.
The post this came from
Watch the reel
All four commands on one card, with the rule and the tool list. Posted on the BANI Academy page.
Reference and copy are not the same thing
Step zero of this method is to find a thumbnail already winning in your niche – from YouTube search rather than an image search – and hand it to the tool as a reference. That is genuinely how most good thumbnails get made, and it is also where people cross a line without meaning to.
| From a reference you like | Verdict | What that means |
|---|---|---|
| Layout and eye path | Borrow freely | Where the subject sits, where the text goes, what the eye hits first |
| Contrast and colour logic | Borrow freely | Bright subject on a dark field; one accent colour doing the work |
| The idea of the shot | Borrow freely | A reaction face beside a clear object |
| Their photo or artwork | Do not reuse | Their face, their stills, their illustration |
| Their text and branding | Do not reuse | Their wording, logo, channel colours or name |
| The whole image, lightly changed | Do not publish | If a viewer would mistake it for theirs, it is not yours |
The practical test is the one in the last row: if someone scrolling past could mistake your thumbnail for the original channel’s, you have copied rather than learned. Change the subject, the text and the colours until it reads as yours – which the card says outright, and which is worth repeating because it is the part that gets skipped.
There is a better version of this habit anyway. Instead of feeding one thumbnail in, study five that work in your niche and write down what they share: how big the face is, how many words are on screen, which colour repeats. That list is a design brief you own, and it is the same reading-the-competitor discipline as surface topic, real audience. And whatever the method, a thumbnail still has to promise what the video delivers – the curiosity-gap rules in thumbnail ideas and the curiosity gap.
Where to run the commands
The card names eight tools, each doing one job in the chain: a reference-editing image tool for the main passes, a chat assistant for text and retouching, a multi-image model when you are combining two pictures, a style-reference model for look, a text-on-image tool for the words, an upscaler for the final size, a cut-out tool for isolating the subject, and a design app for the last polish.
You do not need eight logins. One tool that edits from a reference does passes one to four; everything else is a specialist you add when a particular pass keeps failing – the same one-row-at-a-time rule as generalist versus expert tools. Free tiers on image tools are usually metered per generation, so the grid command has a second benefit: four options for the price of one, which matters when the limits are as tight as those in the limit is the price.
One more habit worth keeping: when a thumbnail shows something that never happened, say so at upload. The reasoning, and how to do it without sounding defensive, is in disclosing generated visuals.
Frequently asked questions
Why four variants and not ten?
Because you can compare four at a glance. Ten becomes another decision to put off, and the grid is meant to end the guessing, not extend it.
What if the face comes out wrong?
Run the subject command again on its own, with “don’t touch the background” kept in. Face swaps fail most often when they are bundled with other edits.
Should I tell viewers the image was generated?
If it shows a realistic scene or person that did not exist, yes – and YouTube asks for that disclosure at upload.
Do I need all eight tools?
No. One image tool that edits from a reference covers most of this. Add an upscaler or a cut-out tool only when a specific pass keeps failing.
Run the four passes on your next video
Find a thumbnail in your niche that is clearly working. Note its layout, not its content. Run the size command, then the edits command, then the subject command with your own photo. Finish with the grid: four variants, one thing changed, numbered.
Then pick a number. That is the difference between guessing and choosing – and it takes four short lines rather than one long one.
Learning a lot does not make anyone good. Doing a lot does.
If you want the structure that turns that into a habit, that is what our free challenge is built for: no barrier to entry, a genuine commitment to action, and a refundable commitment fee you get back when you finish. You are not paying for the knowledge. You are betting on yourself finishing, and we hold the stake.
Want the full set of commands? Comment “BANI” on the original reel and we will send it over.
For more on building channels for English-speaking audiences, that is what we work on at mmoyoutube.com.
