basic flow

• text goes in through a frozen text encoder

• gets turned into something the model can understand (embeddings)

• diffusion model starts with noise and removes it bit by bit

• super-resolution models clean it up and make it bigger

• final image comes out

stack

• text-to-image models (like Imagen, Parti, Stable Diffusion)

• CLIP for understanding what words mean visually

• transformer architecture handling the heavy lifting

• upscalers making small images big

• VQ-GAN for handling the image parts

resources

• Jay Alammar's blog - visual explanations that actually make sense

• Stanford's CS231n course - fundamentals of computer vision

• Hugging Face diffusion course - hands-on with actual models

• AssemblyAI YouTube series on diffusion models

• Andrej Karpathy's neural nets course

• Keras examples of implementing basic models

• Papers:

  • Imagen paper (Google)

  • Parti paper (scaling study)

  • Stable Diffusion paper (for the open source angle)