improve anime
Improve Fooocus Anime a bit by using better SD1.5 refining formulation.
This commit is contained in:
lllyasviel
parent
60c05342b2
commit
736a5aa3ac
@ -69,7 +69,7 @@ vae_approx_filename = os.path.join(vae_approx_path, 'xl-to-v1_interposer-v3.1.sa
|
||||
def parse(x):
|
||||
global vae_approx_model
|
||||
|
||||
x_origin = x['samples'].clone()
|
||||
x_origin = x.clone()
|
||||
|
||||
if vae_approx_model is None:
|
||||
model = Interposer()
|
||||
@ -89,6 +89,5 @@ def parse(x):
|
||||
fcbh.model_management.load_model_gpu(vae_approx_model)
|
||||
|
||||
x = x_origin.to(device=vae_approx_model.load_device, dtype=vae_approx_model.dtype)
|
||||
x = vae_approx_model.model(x)
|
||||
|
||||
return {'samples': x.to(x_origin)}
|
||||
x = vae_approx_model.model(x).to(x_origin)
|
||||
return x
|
||||
|
||||
@ -1 +1 @@
|
||||
version = '2.1.722'
|
||||
version = '2.1.723'
|
||||
|
||||
@ -218,19 +218,21 @@ def get_previewer(model):
|
||||
def ksampler(model, positive, negative, latent, seed=None, steps=30, cfg=7.0, sampler_name='dpmpp_2m_sde_gpu',
|
||||
scheduler='karras', denoise=1.0, disable_noise=False, start_step=None, last_step=None,
|
||||
force_full_denoise=False, callback_function=None, refiner=None, refiner_switch=-1,
|
||||
previewer_start=None, previewer_end=None, sigmas=None, noise=None):
|
||||
previewer_start=None, previewer_end=None, sigmas=None, extra_noise=0.0):
|
||||
|
||||
if sigmas is not None:
|
||||
sigmas = sigmas.clone().to(fcbh.model_management.get_torch_device())
|
||||
|
||||
latent_image = latent["samples"]
|
||||
|
||||
if noise is None:
|
||||
if disable_noise:
|
||||
noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="cpu")
|
||||
else:
|
||||
batch_inds = latent["batch_index"] if "batch_index" in latent else None
|
||||
noise = fcbh.sample.prepare_noise(latent_image, seed, batch_inds)
|
||||
if disable_noise:
|
||||
noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="cpu")
|
||||
else:
|
||||
batch_inds = latent["batch_index"] if "batch_index" in latent else None
|
||||
noise = fcbh.sample.prepare_noise(latent_image, seed, batch_inds)
|
||||
|
||||
if extra_noise > 0.0:
|
||||
noise = noise * (1.0 + extra_noise)
|
||||
|
||||
noise_mask = None
|
||||
if "noise_mask" in latent:
|
||||
|
||||
@ -6,7 +6,7 @@ import modules.path
|
||||
import fcbh.model_management
|
||||
import fcbh.latent_formats
|
||||
import modules.inpaint_worker
|
||||
import modules.sample_hijack as sample_hijack
|
||||
import fooocus_extras.vae_interpose as vae_interpose
|
||||
|
||||
from fcbh.model_base import SDXL, SDXLRefiner
|
||||
from modules.expansion import FooocusExpansion
|
||||
@ -270,22 +270,14 @@ refresh_everything(
|
||||
|
||||
@torch.no_grad()
|
||||
@torch.inference_mode()
|
||||
def vae_parse(x, tiled=False, use_interpose=True):
|
||||
if final_vae is None or final_refiner_vae is None:
|
||||
return x
|
||||
|
||||
if use_interpose:
|
||||
print('VAE interposing ...')
|
||||
import fooocus_extras.vae_interpose
|
||||
x = fooocus_extras.vae_interpose.parse(x)
|
||||
print('VAE interposed ...')
|
||||
def vae_parse(latent, k=1.0):
|
||||
if final_refiner_vae is None:
|
||||
result = latent["samples"]
|
||||
else:
|
||||
print('VAE parsing ...')
|
||||
x = core.decode_vae(vae=final_vae, latent_image=x, tiled=tiled)
|
||||
x = core.encode_vae(vae=final_refiner_vae, pixels=x, tiled=tiled)
|
||||
print('VAE parsed ...')
|
||||
|
||||
return x
|
||||
result = vae_interpose.parse(latent["samples"])
|
||||
if k != 1.0:
|
||||
result = result * k
|
||||
return {'samples': result}
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
@ -444,8 +436,7 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
|
||||
if modules.inpaint_worker.current_task is not None:
|
||||
modules.inpaint_worker.current_task.unswap()
|
||||
|
||||
sample_hijack.history_record = []
|
||||
core.ksampler(
|
||||
sampled_latent = core.ksampler(
|
||||
model=final_unet,
|
||||
positive=positive_cond,
|
||||
negative=negative_cond,
|
||||
@ -467,34 +458,20 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
|
||||
target_model = final_unet
|
||||
print('Use base model to refine itself - this may because of developer mode.')
|
||||
|
||||
# Fooocus' vae parameters
|
||||
k_data = 1.05
|
||||
k_noise = 0.15
|
||||
k_sigmas = 1.4
|
||||
|
||||
sampled_latent = vae_parse(sampled_latent, k=k_data)
|
||||
|
||||
sigmas = calculate_sigmas(sampler=sampler_name,
|
||||
scheduler=scheduler_name,
|
||||
model=target_model.model,
|
||||
steps=steps,
|
||||
denoise=denoise)[switch:]
|
||||
k1 = target_model.model.latent_format.scale_factor
|
||||
k2 = final_unet.model.latent_format.scale_factor
|
||||
k_sigmas = float(k1) / float(k2)
|
||||
sigmas = sigmas * k_sigmas
|
||||
denoise=denoise)[switch:] * k_sigmas
|
||||
len_sigmas = len(sigmas) - 1
|
||||
|
||||
last_step, last_clean_latent, last_noisy_latent = sample_hijack.history_record[-1]
|
||||
last_clean_latent = final_unet.model.process_latent_out(last_clean_latent.cpu().to(torch.float32))
|
||||
last_noisy_latent = final_unet.model.process_latent_out(last_noisy_latent.cpu().to(torch.float32))
|
||||
last_noise = last_noisy_latent - last_clean_latent
|
||||
last_noise = last_noise / last_noise.std()
|
||||
|
||||
noise_mean = torch.mean(last_noise, dim=1, keepdim=True).repeat(1, 4, 1, 1) / k_sigmas
|
||||
|
||||
refiner_noise = torch.normal(
|
||||
mean=noise_mean,
|
||||
std=torch.ones_like(noise_mean),
|
||||
generator=torch.manual_seed(image_seed+1) # Avoid artifacts
|
||||
).to(last_noise)
|
||||
|
||||
sampled_latent = {'samples': last_clean_latent}
|
||||
sampled_latent = vae_parse(sampled_latent)
|
||||
|
||||
if modules.inpaint_worker.current_task is not None:
|
||||
modules.inpaint_worker.current_task.swap()
|
||||
|
||||
@ -504,7 +481,7 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
|
||||
negative=clip_separate(negative_cond, target_model=target_model.model, target_clip=final_clip),
|
||||
latent=sampled_latent,
|
||||
steps=len_sigmas, start_step=0, last_step=len_sigmas, disable_noise=False, force_full_denoise=True,
|
||||
seed=image_seed+2, # Avoid artifacts
|
||||
seed=image_seed,
|
||||
denoise=denoise,
|
||||
callback_function=callback,
|
||||
cfg=cfg_scale,
|
||||
@ -513,7 +490,7 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
|
||||
previewer_start=switch,
|
||||
previewer_end=steps,
|
||||
sigmas=sigmas,
|
||||
noise=refiner_noise
|
||||
extra_noise=k_noise
|
||||
)
|
||||
|
||||
target_model = final_refiner_vae
|
||||
@ -522,5 +499,4 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
|
||||
decoded_latent = core.decode_vae(vae=target_model, latent_image=sampled_latent, tiled=tiled)
|
||||
|
||||
images = core.pytorch_to_numpy(decoded_latent)
|
||||
sample_hijack.history_record = None
|
||||
return images
|
||||
|
||||
@ -11,7 +11,6 @@ from fcbh.samplers import resolve_areas_and_cond_masks, wrap_model, calculate_st
|
||||
|
||||
current_refiner = None
|
||||
refiner_switch_step = -1
|
||||
history_record = None
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
@ -118,9 +117,6 @@ def sample_hacked(model, noise, positive, negative, cfg, device, sampler, sigmas
|
||||
return
|
||||
|
||||
def callback_wrap(step, x0, x, total_steps):
|
||||
global history_record
|
||||
if isinstance(history_record, list):
|
||||
history_record.append((step, x0, x))
|
||||
if step == refiner_switch_step and current_refiner is not None:
|
||||
refiner_switch()
|
||||
if callback is not None:
|
||||
|
||||
@ -1,3 +1,7 @@
|
||||
# 2.1.723
|
||||
|
||||
* Improve Fooocus Anime a bit by using better SD1.5 refining formulation.
|
||||
|
||||
# 2.1.722
|
||||
|
||||
* Now it is possible to translate 100% all texts in the UI.
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user