Why is the flow-DPM Solver so much worse than Euler's method in my flow matching model. FID become very large when use it

[lllzzz3]

@torch.no_grad()
def sample(self, shape):
self.eval()
zt = torch.randn(shape).cuda() ## init the noise

    noise_schedule = NoiseScheduleFlow(schedule='discrete_flow')
    model_fn_continuous = model_wrapper(
                    self.model,
                    noise_schedule,
                    model_type="flow",
                    model_kwargs={},
                    guidance_type="uncond" ,
                    condition= None,
                    classifier_kwargs={},
                    interval_guidance = [0, 1.0],
                )
    dpm_solver = DPM_Solver(
                    model_fn_continuous,
                    noise_schedule,
                    algorithm_type="dpmsolver++",
                )
    zt = dpm_solver.sample(
                    zt, steps=20, order=2,
                    skip_type="time_uniform_flow",
                    method="multistep", 
                    lower_order_final=False,
                    denoise_to_zero=True,
                    solver_type="dpmsolver", 
                    atol=0.0078, rtol=0.05, flow_shift=3.0
        )
    return zt 

[lawrence-cj]

What's your flow matching model's schedule? Could you provide more info about it?

[lllzzz3]

My flow matching model's schedule is 'discrete_flow'. I've fixed this issue, but I've found a new problem, I've found that order=2 works really well most of the time in the case of multi-step solving, but it gets worse with order=3

[lawrence-cj]

Got it. Actually, I haven't tested about order=3 cases. We only use order=2 in our experiments. I'll look into it later, also if you have any insight on it, let me know.