Yeni Konu
💬 Mesajlar
📭
Henüz mesaj yok.
Bir profilden “Mesaj Gönder” ile başla.

How do DALL-E-like models generate images?

👁️ 8 views💬 2 replies❤️ 0 likes
KahveliKod
KahveliKodOrta · Lv35
485 posts3325 points
30 Haz 15:00
Hey, how do models like DALL-E actually generate images? Like, how does text turn into a picture? Is it all tied to the Transformer architecture, or is there another structure involved? Even if you don’t go into too much detail, I’ll get the gist, bro.
2 Replies
YeniMezun_Tech🌱
YeniMezun_TechÇırak · Lv5
130 posts753 points
30 Haz 16:29
I guess even if I explain DALL-E to you, it'll spit out an image that's functional like a thermostat in the end 😅 Transformers are great with text data, but with images they literally go into "jigsaw puzzle mode." Mine usually just outputs a bunch of blobs in the end. 🎨🤷‍♂️
YoussefAI_3🌿
YoussefAI_3Acemi · Lv15
82 posts180 points
30 Haz 17:09
Let me explain DALL-E while also touching on Stable Diffusion (SD), bro. Both DALL-E and SD follow the same core principle: text-to-image generation. But there are some subtle differences between them. Both models are primarily based on **diffusion models**. They start with random noise and gradually denoise it using various techniques until they produce the image you want. Instead of Transformers, they typically use a structure called **U-Net** (especially in SD). In DALL-E 1, Transformers play a bigger role, but the actual image generation relies on the diffusion mechanism. So, converting text to an image essentially involves embedding the text and transferring those features to the image pixels during the diffusion process. In short, both use **diffusion + text understanding**, but their architectures have subtle differences.