hawkes.aiKey ideas
- The one idea — Image generators learn to remove noise, then denoise pure static into a picture step by step — generation as carving, not painting.
- How it is shown — Real image + progressive noise → the reverse learned; then pure static denoised, step by step, into a coherent picture.
- The trap to avoid — Imagining the model "paints" a picture it planned — it removes noise it predicts shouldn't be there, discovering the image as it clears.
- What it sets up — Doing this on millions of pixels is brutally expensive.
AI image generators are trained on one weird job: un-ruining photos. Add noise, learn to remove it — then hand them pure static and watch.
The one idea
Image generators learn to remove noise, then denoise pure static into a picture step by step — generation as carving, not painting.
Everything you know about generation — one token at a time, left to right — throw out for images. Image models work the opposite way: start with pure static and remove noise, step by step, until a picture no one drew emerges. They don't paint. They carve. This is diffusion. How do you learn to carve? Backward. Take millions of photos and destroy them — add noise until each dissolves into static. Destruction is trivial. The model learns the reverse: look at a noisy image, predict what noise to peel away toward the original. Across millions of examples, it becomes an expert at removing chaos. Now the trick.
How it works — the demo
Real image + progressive noise → the reverse learned; then pure static denoised, step by step, into a coherent picture.
Hand the model a fresh field of static that was never a photo. Ask it to denoise anyway. It obeys — but because it learned what real images look like, its removals impose order that was never there. A face resolves. A landscape appears. Creation is confident denoising of nothing. And it's gradual — not one leap but many gentle passes, often twenty to fifty, each removing a little more noise. Early steps settle composition; late steps resolve fine detail. That's the steps setting: more passes, more refinement, more time. The opposite of text's one-word-at-a-time march. Why does this suit images? An image is a whole — lighting, composition, every region depends on every other.
The trap to avoid
Imagining the model "paints" a picture it planned — it removes noise it predicts shouldn't be there, discovering the image as it clears.
Why it matters — and what’s next
Doing this on millions of pixels is brutally expensive.
Diffusion refines the entire canvas together, keeping global coherence instead of committing to one corner before seeing the rest. Its all-at-once refinement fits images the way autoregression fits language. The trap: imagining the model plans the picture, like an artist. There's no plan, no intent — just a denoiser removing what it thinks is noise, and a picture emerging as a side effect of what it learned images look like. Which is why, later, it nails a scene's vibe while botching a hand. It was un-noising, not reasoning about fingers. But carving from static has a brutal cost: every pass chews through the full image, millions of pixels, over and over. Early models needed serious hardware for one picture, keeping image AI in datacenters. The breakthrough was obscenely simple: don't denoise the pixels — denoise a compressed sketch, then decode. Next: latent diffusion.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:45 of narration
Everything you know about generation — one token at a time, left to right — throw out for images. Image models work the opposite way: start with pure static and remove noise, step by step, until a picture no one drew emerges. They don't paint. They carve. This is diffusion.
How do you learn to carve? Backward. Take millions of photos and destroy them — add noise until each dissolves into static. Destruction is trivial. The model learns the reverse: look at a noisy image, predict what noise to peel away toward the original. Across millions of examples, it becomes an expert at removing chaos.
Now the trick. Hand the model a fresh field of static that was never a photo. Ask it to denoise anyway. It obeys — but because it learned what real images look like, its removals impose order that was never there. A face resolves. A landscape appears. Creation is confident denoising of nothing.
And it's gradual — not one leap but many gentle passes, often twenty to fifty, each removing a little more noise. Early steps settle composition; late steps resolve fine detail. That's the steps setting: more passes, more refinement, more time. The opposite of text's one-word-at-a-time march.
Why does this suit images? An image is a whole — lighting, composition, every region depends on every other. Diffusion refines the entire canvas together, keeping global coherence instead of committing to one corner before seeing the rest. Its all-at-once refinement fits images the way autoregression fits language.
The trap: imagining the model plans the picture, like an artist. There's no plan, no intent — just a denoiser removing what it thinks is noise, and a picture emerging as a side effect of what it learned images look like. Which is why, later, it nails a scene's vibe while botching a hand. It was un-noising, not reasoning about fingers.
But carving from static has a brutal cost: every pass chews through the full image, millions of pixels, over and over. Early models needed serious hardware for one picture, keeping image AI in datacenters. The breakthrough was obscenely simple: don't denoise the pixels — denoise a compressed sketch, then decode. Next: latent diffusion.