[PPDiffusers] add DPMSolverMultistepScheduler

[JunnYu]

PR types

New features

PR changes

Others

Description

https://mp.weixin.qq.com/s/vr5Pw6rc36PwQbP7j9vQYg

DPMSolverMultistepScheduler对齐

from ppdiffusers import DPMSolverMultistepScheduler as PPDPMSolverMultistepScheduler
from diffusers import DPMSolverMultistepScheduler as HFDPMSolverMultistepScheduler
import torch
import paddle
import random

init_kwargs = dict(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear")
timestep = 1000
device = "cuda"
batch_size = 16

ppeuler = PPDPMSolverMultistepScheduler(**init_kwargs)
hfeuler = HFDPMSolverMultistepScheduler(**init_kwargs)
def compare(a, b):
    dif = a.cpu().numpy() - b.cpu().numpy()
    meandif, maxdif = abs(dif).mean(), abs(dif).max()
    print(meandif, maxdif)

# test set_timesteps
ppeuler.set_timesteps(timestep)
hfeuler.set_timesteps(timestep)
compare(ppeuler.timesteps, hfeuler.timesteps)

# test add_noise
hflatents = torch.randn((batch_size, 4, 64, 64)).to(device)
hfnoise = torch.randn((batch_size, 4, 64, 64)).to(device)
hftimesteps = torch.randint(0, timestep, (batch_size, )).long().to(device)
pplatents = paddle.to_tensor(hflatents.cpu().numpy())
ppnoise = paddle.to_tensor(hfnoise.cpu().numpy())
pptimesteps = paddle.to_tensor(hftimesteps.cpu().numpy())

hfnoisy_latents = hfeuler.add_noise(hflatents, hfnoise, hftimesteps)
ppnoisy_latents = ppeuler.add_noise(pplatents, ppnoise, pptimesteps)

compare(hfnoisy_latents, ppnoisy_latents)

t = random.randint(0, 1000)
hft = torch.tensor(t, device=device)
ppt = paddle.to_tensor(t)
# test scale_model_input
hfscale_model_input = hfeuler.scale_model_input(hflatents, hft)
ppscale_model_input = ppeuler.scale_model_input(pplatents, ppt)

compare(hfscale_model_input, ppscale_model_input)

# test step
seed = random.randint(0, 2**32)
generator = torch.Generator(device).manual_seed(seed)
paddle.seed(seed)
hfstep = hfeuler.step(hfnoise, hft, hflatents)
ppstep = ppeuler.step(ppnoise, ppt, pplatents)
compare(hfstep[0], ppstep[0])
# 0.0 0.0
# 0.0 0.0
# 4.1596826e-07 7.6293945e-06
# 0.0 0.0
# 8.9845884e-08 7.1525574e-07

import torch
torch.set_grad_enabled(False)
from diffusers import StableDiffusionPipeline as HFStableDiffusionPipeline
import paddle
paddle.set_device("gpu")
paddle.set_grad_enabled(False)
from ppdiffusers import StableDiffusionPipeline as PPStableDiffusionPipeline
device = "cuda"
from ppdiffusers import DPMSolverMultistepScheduler as PPDPMSolverMultistepScheduler
from diffusers import DPMSolverMultistepScheduler as HFDPMSolverMultistepScheduler

hf_scheduler = HFDPMSolverMultistepScheduler.from_config(
    "CompVis/stable-diffusion-v1-4",  # or use the v1-5 version
    subfolder="scheduler",
    solver_order=2,
    predict_epsilon=True,
    thresholding=False,
    algorithm_type="dpmsolver++",
    solver_type="midpoint",
    denoise_final=True,  # the influence of this trick is effective for small (e.g. <=10) steps
)
pp_scheduler = PPDPMSolverMultistepScheduler.from_config(
    "CompVis/stable-diffusion-v1-4",  # or use the v1-5 version
    subfolder="scheduler",
    solver_order=2,
    predict_epsilon=True,
    thresholding=False,
    algorithm_type="dpmsolver++",
    solver_type="midpoint",
    denoise_final=True,  # the influence of this trick is effective for small (e.g. <=10) steps
)
hf_pipe = HFStableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", scheduler=hf_scheduler).to(device)
pp_pipe = PPStableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", scheduler=pp_scheduler)

seed = 42
generator = torch.Generator(device).manual_seed(seed)
img = hf_pipe("a pig", generator=generator, num_inference_steps=20)[0][0]
img.save("hf.png")

seed = 42
img = pp_pipe("a pig", seed=seed, num_inference_steps=20)[0][0]
img.save("pp.png")

[westfish]

可不可以把DPMSolverMultistepScheduler的使用方法案例加入ppdiffusers，最好能有和StableDiffusion使用的普通PNDMScheduler对比的效果，或和其他scheduler对比的效果，我看这里只有和hf DPMSolverMultistepScheduler对比的效果