The ODE in this paper (arxiv.org/abs/2206.00364) lets black box ODE solvers work well with diffusion model ODEs! I tried torchdiffeq's adaptive_heun method and got this result with 38 fevals:

Jun 16, 2022 · 2:53 PM UTC

It's from my cc12m_1_cfg model, and the prompt was "New York City, oil on canvas".
Replying to @RiversHaveWings
Congratulations! Amazing results. And thanks for the paper recommendation.
Replying to @RiversHaveWings
To me, what is more amazing is how they pull an ODE as a substitute for an SDE. That’s a neat trick, justified by the (completely valid) “because it works” argument.
Replying to @RiversHaveWings
Look like what I get using JAX
Replying to @RiversHaveWings
Woah. This looks super effective.