Generative Image / Text-to-Image
In Malay: Imej Generatif
Generative image AI (often called text-to-image) is technology that creates brand-new pictures from a written description or reference image you provide. Rather than cutting and pasting existing photos from a search engine, the model generates a fresh image pixel by pixel based on the subject, style, and lighting in your prompt.
Example: "Portrait of a Malay man at a kopitiam (local coffee shop) at night, warm light" creates a picture that never existed before.
How does AI turn a sentence into a picture?
Most modern image generators rely on a technique called diffusion. During training on millions of paired images and captions, the model learns what visual patterns correspond to words like "warm kopitiam lighting", "wet asphalt after rain", or "batik fabric texture".
When you type a prompt to create a new picture, the model starts with a canvas of pure random static (like an untuned television screen). Step by step, it refines that noise into clear shapes, textures, and lighting that match your written description.
Why generative images are the foundation of great AI video
A common beginner mistake in AI filmmaking is jumping straight into text-to-video and wondering why the main character's face changes in every shot. In my own production workflow, the rule is simple: lock the still image first, then animate it into video.
Generating your still frames first gives you three big advantages before rendering motion:
- You can dial in the exact face, outfit, and Malaysian setting quickly before spending video generation credits.
- That finished picture becomes your starting frame (image-to-video) or character reference inside tools like Higgsfield, Seedance, or Kling.
- Your character consistency across multiple scenes stays believable from the first shot to the final cut.
4 details to include when prompting for AI images
Instead of writing vague adjectives like "cool" or "nice", build your image prompt around four concrete camera details:
- Main subject: Who or what is in frame (for example, a Malay woman in her 30s wearing a navy blazer, or a plate of char kuey teow on a stainless steel table).
- Setting and atmosphere: Where the scene takes place (such as a busy night market right after evening rain).
- Lighting and mood: Soft morning window light, warm tungsten streetlamps, or crisp studio lighting.
- Camera angle and style: Close-up portrait, wide cinematic shot, photorealistic film still, or clean 3D illustration.
Common flaws and copyright considerations
Even the best image models still occasionally struggle with small hands holding objects and long strings of Malay text on shop signs. Whenever you need clean typography on a poster or thumbnail, generate the artwork without text first and add the lettering in your editing app afterward.
On the ethics and copyright side, avoid generating likenesses of real people without their consent, and check each platform's terms before using generated images in commercial campaigns.
Common questions
Does generative image AI just stitch together photos from Google?
No. Diffusion models do not keep a folder of source photos to cut and paste together. They learn mathematical patterns of light, shape, and style during training, then synthesise a new image from random noise each time you prompt them.
Why do AI images often misspell words on signs?
Image models treat letters as visual shapes rather than spelled language the way an LLM does. While newer models handle short words much better, longer Malay phrases still look cleanest when added in an editing app after generating the image.
How do I keep the same character face across multiple AI pictures?
Generate one clean portrait of your character first, then feed that picture back into the tool as an image reference for future scenes while keeping the clothing and facial description consistent in your prompt.
Want to see how still reference images turn into finished AI short films? Browse the recipes at /resipi/ (your first recipe is free with an email) or start at /mula/.