|
|
|
|
|
|
by swyx
386 days ago
|
|
|
> That last point indicates an interesting capability that diffusion models have: you get a kind of built-in quality knob. If you want fast inference at the cost of quality, you can just run the model for less time and end up with more noise in the final output2. If you want high quality and you’re happy to take your time getting there, you can keep running the model until it’s finished removing noise. not quite right... anyone who has run models for >100 steps knows that you can go too far. whts the explanation of that? |
|
|