AI does not really draw
We imagine an artist sitting inside, listening to the description and moving a brush. That is not how it works.
The model is taught with hundreds of millions of images and the descriptions written beside them. During training, noise and blur are added to each image step by step, until the whole thing looks like TV static. The model learns to do the reverse: to bring the picture back out of the static.
When you use it, it starts from random static, treats your description as a compass, and removes the blur over thirty or forty steps until a picture comes out.
This has two consequences. One, if you give the same description twice you will get two different images, because the starting static is different each time. Two, the model does not understand the meaning of words. It only knows what kind of picture usually sits next to which words. So it may not know the word "gamchha" well, but it knows "red and white checked cloth".
From this second point comes a working rule that the rest of this post is built on. Describe the look, not the name. Skip the word the model knows poorly and write what the thing looks like, and the result changes. Instead of "panjabi", write "loose, straight-cut long-sleeved shirt, down to the knee", and you will get much closer to what you want.
How that starting static looks is decided by a number, which many tools call the seed. In some tools you can set this number yourself. The practical meaning: when you get an image you like, note the number. Later, with the same number, if you change one word of the description, the rest of the image stays roughly the same and only that one thing changes. In tools that do not show the number you do not have this option, so when you get an image you like, save it right away.
The five parts of a good description
Most people write "make a beautiful picture". That gives the model no direction. Split the description into five parts and the result changes overnight.
| Part | What to say | Example |
| Subject | The main thing or person, clearly | A middle-aged tea seller |
| Scene | Where, and what is around | A narrow lane in Old Dhaka, early morning |
| Light | What kind of light, from which side | Soft golden light, from behind |
| Style | Photo, oil paint, cartoon, sketch | Documentary photography |
| Frame | How much is visible, from which angle | Waist up, at eye level |
Put all five together and you get: "A middle-aged tea seller in a narrow lane in Old Dhaka in the early morning, soft golden light from behind, in the style of documentary photography, waist-up frame, shot at eye level."
Order matters too. Most models give more weight to the words at the start of the description. So write the most important thing first. For a product image, the product comes first and the background after. If you want a mood or a feeling, put that first.
From a weak description to a good one
Say you want an image for a Eid poster of a sweet shop.
Weak: Make a picture of sweets. Result: a generic cake or a foreign dessert that has nothing to do with your shop.
Medium: Make a picture of Bangladeshi sweets. Result: somewhat closer, but the light and the arrangement are messy.
Good: Rosogolla and kalojam arranged on a white porcelain plate on a wooden table, a clay pot beside them, soft window light from above, in the style of food photography, sharp and detailed, shot from above.
The difference between the three is not talent. It is information. The less you say, the more the model assumes on its own, and those assumptions are usually Western.
How the image changes when you change a word
If you know what each added word does, you do not have to waste time guessing.
| Word to add | What changes in the image |
| Close-up | The subject gets bigger, the background almost disappears |
| Wide shot | The whole setting is visible, the subject gets small |
| 85 mm lens, blurred background | The back goes softly out of focus, it looks like a portrait |
| Soft light | Shadows get lighter, skin looks smooth |
| Hard side light | Shadows get darker, a dramatic feel comes in |
| Studio light, white background | Clean, like a catalogue image |
| Shot from above | Good for food or things laid out on a table |
| Low angle | The subject looks big and serious |
| Film camera grain | An old-photo feel |
| Minimal, leave empty space | Creates empty space for text |
The last row is the most useful when you make posters. If you will place Bangla text on the image later, ask for the empty space up front. Otherwise the image fills up and there is no room for the text.
A saree shop in Mymensingh, the right image in four rounds
Let us look at a real sequence. A saree shop in Mymensingh wants a cover image for its Facebook page before Eid. The photo of the actual saree is already taken on a phone, but they want a nicer background.
Round one. They wrote, "Make a beautiful picture of a saree."
What came out: a model wearing a saree, but the saree looked like an Indian lehenga, with a foreign palace behind. It had nothing at all to do with their shop.
Round two. They added the country: "Make a picture of a Bangladeshi saree."
The result was a bit better, but the light was harsh like a flash, the colours were loud, and the weave of the cloth could not be seen.
Round three. This time they wrote out the five parts: "A folded cotton saree laid out in front of an old wooden almirah, the weave of the cloth clearly visible, a small brass bowl beside it, soft window light from the right, empty space on the left of the image for text, in the style of product photography, shot from slightly above at an angle. There will be no text in the image."
Now the image was usable. Only the saree design was not the shop's.
Round four, and this is the real work. They kept the AI image as the background, placed their own photo of the saree on top in the canvas, then wrote the shop name and the Eid message in the empty space on the left, in a Unicode Bangla font.
What they learned. AI can make a scene. It cannot make your product. The whole thing took forty minutes, of which thirty went on the failed first two rounds. Next time they need the same style of image, the round-three description is saved, and only the item name needs to change. A good description, once written, is an asset. You do not have to think it through again every time.
Eight steps for writing a description
When you need an image, go in this order and you will have fewer random attempts.
- Decide where the image will go. A Facebook post, a page cover, a shop banner and a website header each have a different shape. If you do not fix the shape first, cropping later spoils the image.
- Write the subject in one sentence. Only the main thing, nothing else.
- Add the scene. Where it is, and what is around it.
- Add the light. What kind of light and from which side. This is what makes an image look most professional or most amateur.
- Add the style. Photograph, watercolour, pencil sketch, flat vector design. Name one, do not mix three.
- Add the frame. How much is visible, from which angle, and whether you need empty space for text.
- List what you do not want. No text, no watermark, no extra hands, no crowd.
- Change one thing at a time and run it again. Do not rewrite the whole description. Change the light and see the result. Then change the angle. If you change everything at once, you can no longer tell which word did what.
What to add to change the image
- Camera language. Close-up, wide shot, 85 mm lens, blurred background.
- Time and weather. Late afternoon low sun, cloudy monsoon sky, neon light at night.
- Colour instructions. Red and green dominant, the rest light.
- Material names. Cotton, muslin, brass, bamboo, clay, wood. Naming a material brings a feeling of touch into the image.
- A list of what you do not want. Many tools have a separate box for this. Write what to leave out, for example text, watermark, extra hands.
And the most useful habit is not to expect a perfect image in one go. Pick the one closest to what you wanted from the first round and change one thing at a time in the description. What you get in three or four rounds you would never have got the first time.
Where AI images are still bad
Nobody says this part in the ads, but you will notice it on your first day of real work.
| What you asked for | What you usually get | What to do |
| Bangla text inside the image | Broken letters, wrong conjuncts, upside-down matra | Keep the image empty and add the text later |
| Hands and fingers | Sometimes six fingers, bent wrists | Keep hands out of the frame, or avoid close-ups |
| A specific person's face | It does not match, and there is legal risk too | Take a real photo, do not make it with AI |
| An exact count | Ask for five people, get four or six | Reduce the number, or edit it later |
| Local details | A rickshaw looks like an auto from another country | Give detail in the description, give a sample image |
| Logos and brands | Close but wrong | Place the real logo afterwards |
The second row from the bottom is the most annoying for a Bangladeshi user. The training images had comparatively few of our roads, clothes and shops, so if you ask for "a Dhaka street" you often get a city from another country. Be specific in the description: a three-wheeled pedal rickshaw, a coloured hood, hand-painted designs on the back.
Things image AI still cannot do
The table above is a list of annoyances. This part is more serious, because what is here cannot be fixed by trying again.
The exact look of a specific product. AI cannot make the exact rechargeable fan your shop sells. It will make "something like a fan", with a different switch, wiring and basket design. The model number, the position of the buttons, the number of wires: none of it can be relied on.
Left and right, front and back. The model does not handle spatial relations well. If you say "a bag in the man's left hand, an umbrella in his right", half the time it comes out reversed.
The same character again and again. If a storybook needs six pictures of the same girl, you will get six different girls. Some tools let you keep a face using a sample image, but holding clothes, age and face together is still hard.
Technical accuracy. Machine parts, electrical circuits, organs inside the body, maps, architectural drawings. They will look believable, but the relations inside will be wrong. Using these images for teaching means teaching wrong things.
A specific brand colour. If your organisation's colour follows a fixed code, AI cannot match it. It gives a nearby colour, which shows the moment you put them side by side.
Continuity. If you ask for a before and an after image of the same scene, the furniture, the walls and the clothes will change between the two.
Not one of these six gets fixed by writing a better description. They need another route: a camera, a designer or an editing tool.
Which route to take when you need an image
There are four ways to get an image, not one. Which one to pick depends on what the image will be used for.
| The image you need | Best route | Why |
| Your own shop's product | Shoot it yourself on your phone | Without an exact match, selling does not last |
| A poster background | AI | It need not be specific, and it is fast |
| Decoration for a blog or article | AI | Getting the idea across is enough |
| An organisation's logo | A designer | You need to be the sole owner |
| A familiar place or an event | A real photo | A made-up image deceives the viewer |
| A general scene, such as an office or the sky | AI or a stock photo library | Both work, take the faster one |
| A large printed banner | A camera or a vector design | Much higher resolution is needed |
| The face of a staff member or a customer | A real photo, with permission | A made-up face is false testimony |
Looking at the list, a pattern shows. Where the image has to match a specific reality, do not use AI. Where the image is only there to be pleasing to the eye, AI is the cheapest and fastest route.
If the background of your own product photo is messy, there is a middle route. In the BanglaCodes chat, add the image and press "Remove the background of the image". The item stays exactly as it is and only the background goes, so the rules above are not broken. The button next to it, "Give a new background", does redraw the image with AI and may change small details of the item, so for a selling photo the first one is the safe one. Both are free.
If the image needs to move
The whole table above is about still images. But in comments, on WhatsApp or in email, movement often works better, and a video is too much there. A GIF of a few seconds is then the easiest route: the file is small, it goes through on a weak network, nobody has to press play, and it loops by itself.
If you ask for a transparent background, the badge sits on a post of any colour. Keep it to three to six seconds and two or three colours and the file stays light. For anything longer than that, video is better.
The right system for putting Bangla text on an image
This needs a separate mention, because this is the one place where the most time gets wasted.
Bangla script has conjuncts, the matra, and the same sound can be written in more than one way. An image model does not recognise letters as letters. It sees shapes. So Bangla text almost always comes out broken: it looks like Bangla but cannot be read. English has got better lately, but Bangla is still not something to rely on.
The working system is these three steps.
- Use AI only for the background or the scene. Write in the description that there should be no text, and ask it to leave empty space for the text.
- Download the image and open it in Canva, Figma or Photoshop.
- Place the text yourself in a Unicode Bangla font.
This way the spelling stays in your control, the font is the one you like, and when a price or a date changes later you do not have to make the whole image again.
One extra tip: build the text template once and save it. Each month you only change the background and the date, and a new poster is ready. This also keeps your shop looking the same, which matters for a brand.
Mistakes that keep happening, and the fix
| Mistake | What happens | Fix |
| Asking for ten things in one description | The model drops some of them | Keep one main subject, do the rest later in editing |
| Rewriting the whole description each time | You cannot tell which word did what | Change one thing at a time |
| Words like "beautiful", "wonderful" | No direction, the result is random | Describe what it should look like |
| Starting without fixing the image shape | Cropping later cuts off the main thing | Fix the space first, then the image |
| Asking for Bangla text inside the image | Broken letters, unreadable | Keep the image empty and add the text later |
| Using the first image you get | Quality stays average | Make three or four and take the best |
| Making a banner at low resolution | It looks blurry in print | Make it in a large size from the start |
Copyright and ownership
You hear two kinds of exaggeration about this: "everything is free, do what you like" and "everything is illegal". The reality is in the middle, and it is not settled yet.
Do not mix up three separate questions.
One, whether you can use the image commercially. This is decided by the terms of the tool you use. Many tools allow commercial use on a paid plan, and on the free plan they do not, or they add a watermark. Read the terms once before you use an image for business.
Two, whether you get the copyright to the image. Whether a person holds copyright in a fully AI-made image is under discussion in country after country, and in many places it has not been clearly settled. In practical terms, if someone else uses your AI image, it may be hard to stop them. Do not leave things that must be yours alone, such as a brand logo, to AI.
Three, whether the image infringes someone's rights. This is where the risk is most real. I am not a lawyer, and this post is not legal advice. Making and using someone else's logo, a well-known cartoon character, or the likeness of a living person is risky. Changing someone's photo without permission and spreading it is not only a legal risk, it is also wrong.
Where to use it and where not
Use it without worry. Backgrounds for blog posts, drafts to show an idea, design inspiration, social media decoration, internal presentations.
Do not. Putting an AI image in place of the real product photo deceives the buyer, because when the customer gets the item and it does not match, the trust breaks at once. Food photos are risky for the same reason: if your biryani does not look like the picture, you will find out in the reviews.
Do not use it. As a news photo, in medical or legal documents, in a photo for anyone's ID, and anywhere the viewer will assume the image was taken in real life.
A simple test: if the viewer would feel cheated on finding out, do not use an AI image there. And if you do use one, add a short note that the image was made with AI.
When not to use it
The section above was about ethics. This section is about work: which jobs image AI is simply the wrong tool for.
When you have the real thing in your hands. Take an electronics shop in Bogura. The owner thought he would make a shiny picture of a rice cooker with AI, because the phone photo was blurry. The result: the cooker looked nice in the image, but it had three buttons instead of four, and the lid handle was different. The customer got the item and complained. The fix was much cheaper: spread a sheet of white paper beside the window and take the photo again on the phone. When the item is in front of you, the camera is the right tool.
When an exact match is needed. A uniform design, the craftsmanship on a piece of jewellery, a particular artist's style on a book cover, an organisation's colour. In any job where "close" is not good enough, an AI image wastes your time.
When the image is the proof. A compensation claim, an insurance paper, construction progress, a medical record. Using a made-up image here is not just wrong, it is fraud.
When the person is real. Putting an AI face in as a staff member, teacher, doctor or customer, especially as a "satisfied customer". That is manufacturing false testimony.
When you have no time. This sounds odd, but it is true. Getting the image you want usually takes several rounds. If you need one specific image within fifteen minutes, buying one from a stock library or taking it yourself is often faster.
The whole path for ads with images and video is in the post on UGC ads.
If you want to make one yourself, open the AI image page.
Questions and answers
Can you tell an image was made with AI? Often you can. It shows in the fingers, teeth, earrings, text in the background and how a held object connects to the hand. But quality is improving fast, so trusting the naked eye is getting less reliable. Some tools put an invisible mark inside the file.
Is it safe to upload my own photo and edit it? Read the tool's terms. If it is your own photo the risk is yours, but before you upload photos of family or staff, you should get their permission. It is better not to upload photos of children at all.
Is the image resolution enough for print? For social media it is usually enough. For a banner or a billboard it often is not. The free plan gives small sizes, and enlarging needs a separate tool that still loses detail.
Can I make the same character again and again? It is hard. Even with the same description the face changes. Some tools let you hold a character using a sample image, but do not expect a perfect match yet.
Should I write the description in Bangla for image making? Most image models work better with English descriptions, because the descriptions used in training were mainly in English. If you write in Bangla and do not get good results, translate the description into English.
Is it right to change the background of a photo I took myself? For ordinary decoration there is no problem. But in a product selling photo, changing the background to make the item look different, such as making a small thing look big or wiping out a mark, counts as misleading the buyer. Cleaning up is one thing, changing is another. "Remove the background of the image" in the BanglaCodes chat does not touch the item, it only removes the background. "Give a new background" redraws the whole image, so keep that for posters and festival decoration.
How many images should I make before picking the best? For ordinary work, three or four are enough. If you make more, the time goes into choosing. It is better to fix the description and run it again, because twenty images from a weak description will all be weak.
Will image AI take designers' jobs? The pressure on drafts and idea-showing work will go down, that is true. But design decisions, brand consistency, print preparation and talking with the client are not things AI does. In practice many designers now use AI as a tool for drafts.
Is it okay to put someone's face into an image for fun? Not without permission. It may feel like fun among friends, but once the image spreads it is no longer in your hands. And changing someone's photo in an indecent way is never acceptable.
In short
- AI does not draw. From random static it brings out, step by step, an image that matches your description.
- A good description has five parts: subject, scene, light, style and frame. If you change one thing at a time and run it again, the image you want comes quickly.
- Bangla text inside an image almost always comes out broken, so keep the image empty and add the text later in a Unicode font.
- AI is still weak at hands, fingers, exact counts and specific people's faces. And local details such as a rickshaw or a lane in Old Dhaka turn into a scene from another country unless you describe them in detail.
- Permission for commercial use is in the tool's terms, and who holds the copyright of an AI image is still under discussion in country after country.
- Do not use AI images as news photos, in medical documents or in ID photos, and if you do use one, say so.