Tạo hình ảnh AI: Cách hoạt động của mô hình chuyển văn bản thành hình ảnh

An AI image generator is a program that has studied a very large collection of images, usually paired with captions, and learned the visual patterns that connect words to pictures. When given a prompt such as a red fox in snow, watercolor style, it does not search for an existing picture or paste together pieces of stored photos. It produces new pixels that are statistically consistent with what it learned about foxes, snow and watercolor. The output can be a finished image, a variation of an uploaded image, or an edit of one region of a photo.

The most widely used mechanism today is diffusion. During training, the system takes real images and adds random noise to them in many small steps until they become pure static. A neural network learns to reverse one step at a time by predicting the noise that was added. To generate, the process runs backward: start with random static, ask the network to remove a little noise, repeat dozens of times, and an image gradually emerges. A text prompt steers each step so that the emerging image matches the description. Because each run starts from different random noise, called the seed, the same prompt gives different images every time unless the seed is fixed.

Two supporting pieces make text-to-image work. The first is a text encoder that turns the prompt into numbers the image model can use. OpenAI's CLIP, posted in February 2021, learned to match images with captions from 400 million image and text pairs collected from the internet, and models like it became the bridge between language and pictures. Google's Imagen paper in May 2022 found that using a larger text-only language model as the encoder improved image quality and prompt matching more than enlarging the image model itself. The second piece is efficiency. Running diffusion on full-size pixels is slow, so the 2021 latent diffusion paper by Rombach and colleagues, which became the basis of Stable Diffusion, first compresses images with an autoencoder, a network that squeezes an image into a much smaller grid of numbers and can expand it back. Diffusion then runs in that compressed latent space, and the result is decoded into pixels at the end. The paper also added cross-attention layers so the model can be conditioned on text, layouts or other inputs.

Most tools offer several modes built on the same model. Text-to-image creates a picture from a prompt alone. Image-to-image starts from an uploaded picture partly covered in noise, so the result keeps its rough composition while changing style or detail. Inpainting regenerates only a masked region, such as replacing a sky or removing an object, and outpainting extends an image beyond its original borders; the latent diffusion paper reported state-of-the-art inpainting results. Some systems also accept layouts, sketches or bounding boxes that fix where objects should appear, which the same paper's cross-attention design was built to support. Three settings matter in almost every tool. The seed fixes the starting noise so a result can be reproduced. The number of steps trades speed for refinement. The guidance strength controls how literally the model follows the prompt, with high values giving closer adherence but less variety and sometimes harsh, oversaturated images. Upscaling then enlarges the final picture, but it invents plausible detail rather than recovering real detail that was never there.

The field moved quickly. Generative adversarial networks (GANs), introduced by Ian Goodfellow and colleagues in June 2014, pitted a generator network against a discriminator network that tries to spot the fakes. In February 2021 OpenAI described DALL-E, which treated text and image tokens as one long sequence for a transformer to predict, and CLIP appeared the same month. Latent diffusion followed in December 2021. In 2022 came DALL-E 2 in April, Imagen in May and the public release of Stable Diffusion on August 22, which put a capable model on ordinary graphics cards; Stability AI's announcement put its memory use at 6.9 GB. Since then, commercial tools and open models have improved resolution, prompt following and the ability to render text inside images.

It helps to compare the main approaches. GANs generate an image in a single pass, so they are fast, but they are hard to train and tend to cover a narrower range of images. Autoregressive models, such as the first DALL-E, build an image piece by piece like writing a sentence, which is flexible but slow for large images. Diffusion models are slower at generation because they need many denoising steps, but they train stably and produce varied, detailed results, which is why they dominate current tools. Compared with stock photography or human illustration, a generator is fast and cheap per image, but it cannot guarantee that a picture shows anything that really happened, and its legal status is less settled.

The limitations are concrete. Models often struggle with exact counts, precise spatial relationships, readable text and consistent details such as hands or the same character across several images. They reproduce patterns in their training data, which can include stereotypes about who appears in which roles. Because training sets were largely scraped from the web, there are open legal disputes about whether training on copyrighted images without permission is lawful; the US Copyright Office released a pre-publication report on generative AI training in May 2025. Separately, the Office's January 2025 report on copyrightability concluded that prompts alone do not provide enough human control for the output to be protected, although human-authored elements, creative selection and arrangement, and modifications can be. Realistic fakes of real people and events are the most serious misuse, which is why provenance standards such as C2PA Content Credentials record how an image was made.

In practice, there are two ways to use image generation. Hosted services run the model in the cloud and usually charge a subscription or per-image credits, with the provider's content rules and terms of use deciding what is allowed and who may use the outputs commercially. Open models such as Stable Diffusion can be run locally; the original 2022 release needed a graphics card with about 6.9 GB of memory, and its license allowed commercial and non-commercial use subject to use restrictions. Training a new model from scratch is far more expensive: the latent diffusion paper notes that pixel-space diffusion models often needed hundreds of GPU days to train, which is the cost its compressed approach was designed to reduce. For most people the real work is iteration: writing clear prompts, generating several candidates, editing the best one, checking it at final size, and labeling it honestly as AI-generated.