Same velocity field as above: data at Β±ΞΌ, noise π©(0,1), marginal velocity in closed form.
Both solvers integrate dx/dt = v(x,t) backward from t=1 (noise) to t=0 (data).
Euler uses one velocity evaluation per step (= DDIM).
Heun uses two evaluations with predictβcorrect (= DPM-Solver-2).
Use Step βΆ to watch each method work, one step at a time.