Mirrored_Repos/InvokeAI
1
0
Fork 0
mirror of https://github.com/invoke-ai/InvokeAI synced 2024-08-30 20:32:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'switch-ksampler-noise-scheduler-adaptively' into development

Browse Source 
- This sets a step switchover point at which the k-samplers stop using the
  Karras noise schedule and start using the LatentDiffusion noise schedule.
  The advantage of this is that the Karras schedule produces excellent
  results at low step counts but starts to become unstable at high
  steps.

- A new command argument --karras_max, lets the user set where the
  switchover occurs. Default is 29 steps (1-29 steps Karras),
  (30 or greater LDM)

- Tildebyte, sorry to do a fast forward three-way merge for this
  but rebasing was just too painful due to extensive recent
  changes to the diffuser code.
This commit is contained in:
Lincoln Stein 2022-10-27 16:11:26 -04:00
commit 943616044a
5 changed files with 52 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
docs/features/CLI.md
View File

@ -153,6 +153,7 @@ Here are the invoke> command that apply to txt2img:
| --cfg_scale <float>| -C<float> | 7.5 | How hard to try to match the prompt to the generated image; any number greater than 1.0 works, but the useful range is roughly 5.0 to 20.0 |
| --seed <int> | -S<int> | None | Set the random seed for the next series of images. This can be used to recreate an image generated previously.|
| --sampler <sampler>| -A<sampler>| k_lms | Sampler to use. Use -h to get list of available samplers. |
| --karras_max <int> | | 29 | When using k_* samplers, set the maximum number of steps before shifting from using the Karras noise schedule (good for low step counts) to the LatentDiffusion noise schedule (good for high step counts) This value is sticky. [29] |
| --hires_fix | | | Larger images often have duplication artefacts. This option suppresses duplicates by generating the image at low res, and then using img2img to increase the resolution |
| --png_compression <0-9> | -z<0-9> | 6 | Select level of compression for output files, from 0 (no compression) to 9 (max compression) |
| --grid | -g | False | Turn on grid mode to return a single image combining all the images generated by this prompt |

27
ldm/generate.py
View File

@ -180,6 +180,7 @@ class Generate:
self.size_matters = True # used to warn once about large image sizes and VRAM
self.txt2mask = None
self.safety_checker = None
self.karras_max = None
# Note that in previous versions, there was an option to pass the
# device to Generate(). However the device was then ignored, so
@ -270,6 +271,7 @@ class Generate:
variation_amount = 0.0,
threshold = 0.0,
perlin = 0.0,
karras_max = None,
# these are specific to img2img and inpaint
init_img = None,
init_mask = None,
@ -353,7 +355,8 @@ class Generate:
strength = strength or self.strength
self.seed = seed
self.log_tokenization = log_tokenization
self.step_callback = step_callback
self.step_callback = step_callback
self.karras_max = karras_max
with_variations = [] if with_variations is None else with_variations
# will instantiate the model or return it from cache
@ -398,6 +401,11 @@ class Generate:
self.sampler_name = sampler_name
self._set_sampler()
# bit of a hack to change the cached sampler's karras threshold to
# whatever the user asked for
if karras_max is not None and isinstance(self.sampler,KSampler):
self.sampler.adjust_settings(karras_max=karras_max)
tic = time.time()
if self._has_cuda():
torch.cuda.reset_peak_memory_stats()
@ -878,26 +886,23 @@ class Generate:
# consistent, at least
def _set_sampler(self):
msg = f'>> Setting Sampler to {self.sampler_name}'
karras_max = self.karras_max # set in generate() call
if self.sampler_name == 'plms':
self.sampler = PLMSSampler(self.model, device=self.device)
elif self.sampler_name == 'ddim':
self.sampler = DDIMSampler(self.model, device=self.device)
elif self.sampler_name == 'k_dpm_2_a':
self.sampler = KSampler(
self.model, 'dpm_2_ancestral', device=self.device
)
self.sampler = KSampler(self.model, 'dpm_2_ancestral', device=self.device, karras_max=karras_max)
elif self.sampler_name == 'k_dpm_2':
self.sampler = KSampler(self.model, 'dpm_2', device=self.device)
self.sampler = KSampler(self.model, 'dpm_2', device=self.device, karras_max=karras_max)
elif self.sampler_name == 'k_euler_a':
self.sampler = KSampler(
self.model, 'euler_ancestral', device=self.device
)
self.sampler = KSampler(self.model, 'euler_ancestral', device=self.device, karras_max=karras_max)
elif self.sampler_name == 'k_euler':
self.sampler = KSampler(self.model, 'euler', device=self.device)
self.sampler = KSampler(self.model, 'euler', device=self.device, karras_max=karras_max)
elif self.sampler_name == 'k_heun':
self.sampler = KSampler(self.model, 'heun', device=self.device)
self.sampler = KSampler(self.model, 'heun', device=self.device, karras_max=karras_max)
elif self.sampler_name == 'k_lms':
self.sampler = KSampler(self.model, 'lms', device=self.device)
self.sampler = KSampler(self.model, 'lms', device=self.device, karras_max=karras_max)
else:
msg = f'>> Unsupported Sampler: {self.sampler_name}, Defaulting to plms'
self.sampler = PLMSSampler(self.model, device=self.device)

10
ldm/invoke/args.py
View File

@ -215,6 +215,8 @@ class Args(object):
switches.append(f'-W {a["width"]}')
switches.append(f'-H {a["height"]}')
switches.append(f'-C {a["cfg_scale"]}')
if a['karras_max'] is not None:
switches.append(f'--karras_max {a["karras_max"]}')
if a['perlin'] > 0:
switches.append(f'--perlin {a["perlin"]}')
if a['threshold'] > 0:
@ -691,7 +693,13 @@ class Args(object):
default=6,
choices=range(0,10),
dest='png_compression',
help='level of PNG compression, from 0 (none) to 9 (maximum). Default is 6.'
help='level of PNG compression, from 0 (none) to 9 (maximum). [6]'
)
render_group.add_argument(
'--karras_max',
type=int,
default=None,
help="control the point at which the K* samplers will shift from using the Karras noise schedule (good for low step counts) to the LatentDiffusion noise schedule (good for high step counts). Set to 0 to use LatentDiffusion for all step values, and to a high value (e.g. 1000) to use Karras for all step values. [29]."
)
img2img_group.add_argument(
'-I',

16
ldm/models/diffusion/ksampler.py
View File

@ -8,6 +8,10 @@ from .sampler import Sampler
from .shared_invokeai_diffusion import InvokeAIDiffuserComponent
# at this threshold, the scheduler will stop using the Karras
# noise schedule and start using the model's schedule
STEP_THRESHOLD = 29
def cfg_apply_threshold(result, threshold = 0.0, scale = 0.7):
if threshold <= 0.0:
return result
@ -64,6 +68,9 @@ class KSampler(Sampler):
self.sigmas = None
self.ds = None
self.s_in = None
self.karras_max = kwargs.get('karras_max',STEP_THRESHOLD)
if self.karras_max is None:
self.karras_max = STEP_THRESHOLD
def make_schedule(
self,
@ -92,8 +99,13 @@ class KSampler(Sampler):
rho=7.,
device=self.device,
)
self.sigmas = self.model_sigmas
#self.sigmas = self.karras_sigmas
if ddim_num_steps >= self.karras_max:
print(f'>> Ksampler using model noise schedule (steps > {self.karras_max})')
self.sigmas = self.model_sigmas
else:
print(f'>> Ksampler using karras noise schedule (steps <= {self.karras_max})')
self.sigmas = self.karras_sigmas
# ALERT: We are completely overriding the sample() method in the base class, which
# means that inpainting will not work. To get this to work we need to be able to

15
ldm/models/diffusion/sampler.py
View File

@ -2,10 +2,7 @@
ldm.models.diffusion.sampler
Base class for ldm.models.diffusion.ddim, ldm.models.diffusion.ksampler, etc
'''
from math import ceil
import torch
import numpy as np
from tqdm import tqdm
@ -439,3 +436,15 @@ class Sampler(object):
def uses_inpainting_model(self)->bool:
return self.conditioning_key() in ('hybrid','concat')
def adjust_settings(self,**kwargs):
'''
This is a catch-all method for adjusting any instance variables
after the sampler is instantiated. No type-checking performed
here, so use with care!
'''
for k in kwargs.keys():
try:
setattr(self,k,kwargs[k])
except AttributeError:
print(f'** Warning: attempt to set unknown attribute {k} in sampler of type {type(self)}')