10 steps, first six sigmas and the smallest non-zero one
normal        14.615   7.840   4.609   2.918   1.950   1.345  ... 0.029
simple        14.615   8.303   5.088   3.321   2.276   1.613  ... 0.342
sgm_uniform   14.615   8.307   5.092   3.325   2.280   1.616  ... 0.344
karras        14.615   9.103   5.478   3.169   1.749   0.914  ... 0.029
exponential   14.615   7.325   3.671   1.840   0.922   0.462  ... 0.029
kl_optimal    14.615   4.232   2.394   1.597   1.133   0.815  ... 0.029
beta          14.615  11.543   7.372   4.373   2.615   1.618  ... 0.199

RMS error against a 6000-step reference
schedule       euler@10   euler@20   dpm2m@10   dpm2m@20
normal         0.12316   0.06500   0.23396   0.07024
simple         0.12088   0.06954   0.08455   0.05181
sgm_uniform    0.12118   0.06996   0.08504   0.05251
karras         0.15176   0.07800   0.05459   0.01037
exponential    0.15635   0.08059   0.03931   0.00898
kl_optimal     0.10740   0.05393   0.06005   0.01579
beta           0.12001   0.06399   0.04828   0.01567

flow schedule, 8 steps, sigma = shift(linspace(1, 0))
shift 1.0  1.000  0.875  0.750  0.625  0.500  0.375  0.250  0.125
shift 1.5  1.000  0.913  0.818  0.714  0.600  0.474  0.333  0.176
shift 3.0  1.000  0.955  0.900  0.833  0.750  0.643  0.500  0.300
shift 6.0  1.000  0.977  0.947  0.909  0.857  0.783  0.667  0.462

Euler RMS error on this 1-D toy, against a reference
steps  shift=1.0  shift=1.5  shift=3.0  shift=6.0
4      0.25871    0.26221    0.37637    0.60039  
8      0.14107    0.14446    0.18287    0.30681  
16     0.07621    0.07792    0.09929    0.15182  
32     0.03989    0.04084    0.05266    0.08165  

FLUX mu = base + (max-base) * (tokens-256)/(4096-256), with base_shift 0.5, max_shift 1.15, exp(mu) is the effective shift
  512 px   1024 tokens  mu=0.630  exp(mu)=1.878
  768 px   2304 tokens  mu=0.847  exp(mu)=2.332
 1024 px   4096 tokens  mu=1.150  exp(mu)=3.158
 1536 px   9216 tokens  mu=2.017  exp(mu)=7.513
