I’m a 19 y/o founder from Germany, and I cracked how Google’s Text Diffusion model works! There are three main approaches on this topic:
1. D3PM / SEDD: Puzzle-style Token Repair Mask-and-fix diffusion. Sentences are gradually “shredded” into ⟨MASK⟩ tokens; a transformer learns to un-shred them step-by-step. Imagine ChatGPT playing Wordle, but at 100x speed. (Discrete forward process + cross-entropy
2. Diffusion-LM: Noise-to-Meaning with Classifier Guidance Static-to-speech diffusion. Add Gaussian noise to token embeddings → predict the noise out. Because it’s continuous, you can steer with gradients (sentiment, style, etc.) like turning a radio dial until the song is
3. BlockDiff (Gemini): High-Speed Blockwise Diffusion LEGO-blocks diffusion. Generate 16-token blocks in parallel, then snap them together left-to-right. 8 denoise passes per block, KV-cache reused → 1.5 k token/s on TPU. Same fluency as autoregressive, 5× faster. Think
