Added brownian_noise to DMP 2++ SDE Scheduler and fixed use_exponential_sigmas behaviour

[Crized-bit]

What does this PR do?

Added the option to supply custom Brownian noise generator as was made in AUTO1111 webui.

Also fixed issue with use_exponential_sigmas crashes (nonetype and order issue)

Before submitting

• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@yiyixuxu

[sayakpaul]

@Crized-bit thanks for your PR!

Also fixed issue with use_exponential_sigmas crashes (nonetype and order issue)

Could you maybe supplement a minimal reproducible snippet to understand this better?

[Crized-bit]

Here it is! With this code i had issues that i've fixed

import torch
import cv2
import numpy as np
import os
from tqdm import tqdm
from PIL.Image import fromarray
from diffusers import StableDiffusionXLControlNetPipeline # type: ignore
from diffusers import ControlNetModel, DPMSolverMultistepScheduler # type: ignore
from typing import Callable
from compel import Compel, ReturnedEmbeddingsType
import torch
import torch.nn as nn
def get_compel_embeddings(pipeline, device: torch.device) -> Callable[[str, str], tuple]:
compel=Compel(tokenizer=[pipeline.tokenizer, pipeline.tokenizer_2],
text_encoder=[pipeline.text_encoder, pipeline.text_encoder_2],
returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED, requires_pooled=[False, True],
truncate_long_prompts=False,
device=device) # type: ignore
def return_embeddings(positive_prompt: str, negative_prompt: str):
conditioning, pooled_prompt_embeds=compel(positive_prompt)
negative_conditioning, negative_pooled_prompt_embeds=compel(negative_prompt)
[con_embeds, neg_embeds] = compel.pad_conditioning_tensors_to_same_length([conditioning, negative_conditioning])
return con_embeds, pooled_prompt_embeds, neg_embeds, negative_pooled_prompt_embeds
return return_embeddings
# Default parameters block
DEVICE = torch.device("cuda:2")
output_folder = "/home/Diffusion_VK/results/longClip/"
control_folder = "/home/dataset/Games/"
# ControlNets block
controlnets = ControlNetModel.from_pretrained(
"/home/web_ui/stable-diffusion-webui/models/ControlNet/controlnet_tile", torch_dtype=torch.bfloat16, use_safetensors=True)
# Creating a pipe
pipeline = StableDiffusionXLControlNetPipeline.from_single_file(
# "/home/web_ui/stable-diffusion-webui/models/Stable-diffusion/cyberrealisticXL_v22.safetensors",
"/home/web_ui/stable-diffusion-webui/models/Stable-diffusion/juggernautXL_juggXIByRundiffusion.safetensors",
use_safetensors=True,
torch_dtype=torch.bfloat16, controlnet = controlnets)
pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config,
algorithm_type="sde-dpmsolver++",
final_sigmas_type = "zero", timestep_spacing="linspace", use_exponential_sigmas=True,
)
# Offload to GPU
pipeline.to(DEVICE)
# Create embedder instance to generate embeddings
embedder = get_compel_embeddings(pipeline=pipeline,
device=DEVICE)
# Dataset inference
for filename in tqdm(sorted(os.listdir(control_folder))):
img = cv2.imread(os.path.join(control_folder, filename))
controlnet_input = fromarray(img)
# Generate image
images = []
if os.path.isfile(os.path.join("/home/dataset/Games_captions", filename[:-4] + '.txt')):
with open("/home/dataset/Games_captions/" + filename[:-4] + '.txt', "r", encoding="utf-8") as f:
positive_promt = f.read()
else:
continue
negative_promt = 'bad eyes, cgi, airbrushed, plastic, deformed, watermark, polygons, inaccurate body, worst quality, low quality, normal quality, jpeg artifacts, unrealistic, flat, triangular, low resolution, bad composition'
con_embeds, pooled_prompt_embeds, neg_embeds, negative_pooled_prompt_embeds = embedder(positive_promt, negative_promt)
# Generate image
images = []
# images.append(np.array(controlnet_input))
result_img = pipeline(prompt_embeds=con_embeds,
pooled_prompt_embeds = pooled_prompt_embeds,
negative_prompt_embeds=neg_embeds,
negative_pooled_prompt_embeds = negative_pooled_prompt_embeds,
image=controlnet_input,
generator=torch.Generator(device=DEVICE).manual_seed(1787813600),
num_inference_steps = 30,
controlnet_conditioning_scale = 0.5,
guidance_scale = 5.,
).images[0] # type: ignore
images.append(np.array(result_img))
# Then prepare list of np.arrays' to hcat, then do it
cv2.imwrite(output_folder + filename, cv2.cvtColor(cv2.hconcat(images), cv2.COLOR_RGB2BGR))
```

[yiyixuxu]

was it fixed in #9954?

[Crized-bit]

was it fixed in #9954?

Seems so! So, just added noise

[sayakpaul]

Seems so! So, just added noise

What does this mean?

[Crized-bit]

Seems so! So, just added noise

What does this mean?

This means that the half of my PR was fixed by another PR (#9954) , but the part where i've added brownian_noise (as it's done in k-diffusers and AUTO-1111) is unique and in need of review.

[sayakpaul]

@yiyixuxu could give a look?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.