Text-to-image

Text-to-image models take a natural language prompt ("a cyberpunk cat samurai in neon Tokyo, cinematic lighting") and output a synthesized image. The category exploded in 2022 with Stable Diffusion and DALL-E 2, and by 2026 has fractured into a few clear winners: Midjourney for stylized aesthetics, Flux for photorealism and prompt adherence, Stable Diffusion / SDXL / SD3 as the open ecosystem workhorse, DALL-E (inside ChatGPT) for casual users, and Ideogram for anything involving readable text inside images. Almost all of them work the same way under the hood: diffusion. The model starts with pure noise and gradually denoises it over 20-50 steps, guided by a text encoder (usually CLIP or T5) that turns your prompt into an embedding the image generator can condition on. Most modern systems are *latent* diffusion — they denoise in a compressed latent space rather than at pixel resolution, which is why a 1024x1024 image takes seconds rather than minutes. A few models (like Google's Imagen earlier generations) used pixel-space diffusion or cascaded super-resolution, but latent diffusion won. The differences between tools are mostly about training data, aesthetic bias, and ecosystem. Midjourney is closed and opinionated — its "house style" is baked in and you can't replace it. Stable Diffusion and Flux are open weights, which means a massive ecosystem of LoRAs, ControlNets, and fine-tunes. DALL-E inside ChatGPT is the easiest to access but the least controllable. Picking one is more about workflow than raw quality at this point.