【WebUI 1111】実験用サンプラーRestartLab【Stable diffusion】

BDさんの記事を見て、面白そうだったのでコードを見てたらサンプラー変えれそうだったので実験用サンプラーを作ってみました。

新サンプラーのアイデア

https://note.com/br_d/n/n276380a05ecf

インストール

RestartLab.py をダウンロードして WebUI の Scriptsフォルダーに移動します。

過去のバージョンは、BOOTHにあります。
支援する事も出来ますので、役に立ったと思われた方はよろしくお願いします。
このリンクから→楽天、アマゾンでお買い物でも支援できます。
#RestartLab ver 2.0 2023 yukitoki.blogspot.com #v2.0 2024/1/20 #LCMのノイズを変更できるようにしました。デフォルトはAncestralです。 #UIを少し変更 #sigma churnとsigma noizeをコントロールできるようにしました。 #Auto skipを追加 #Auto skipした途中から生成再開できるようにしました。 #AutoSkip用にBatch処理を追加。 #v1.3 2024/1/16 #xyzでchangeCFGとstepに小数点が使えない不具合を修正 #v1.2 2023/12/4 #PnginfoやSettingで、NoiseがBranian remakeになっているとSDEで処理される不具合を修正しました。 #v1.1 2023/12/3 #samplingにNoizeを追加 #Change CFG Stepを 1にした時 100%にしました。 1stepにしたい時は1.1を指定してください。 #Change CFG Stepの割合の計算が間違ってたので修正しました。 #change CFG modeを追加しました。切替方法を選べるようになりました。 #NoiseのBrownianがBranianになってたので修正しました。 #NoiseのDPM++ SDEをSDEに名前を変更しました。 import torch import io import tqdm import k_diffusion.sampling from modules import sd_samplers_common, sd_samplers_kdiffusion, sd_samplers from k_diffusion.sampling import default_noise_sampler, trange,BrownianTreeNoiseSampler import modules.scripts as scripts import gradio as gr from modules.script_callbacks import CFGDenoisedParams, on_cfg_denoised,ImageSaveParams, on_before_image_saved, AfterCFGCallbackParams, on_cfg_after_cfg, on_infotext_pasted from modules import script_callbacks from modules import shared, sd_samplers_cfg_denoiser from k_diffusion import utils, sampling from k_diffusion.external import DiscreteEpsDDPMDenoiser import base64 import zlib import math import numpy import ast from modules.shared import state import modules.generation_parameters_copypaste as parameters_copypaste from modules.generation_parameters_copypaste import parse_generation_parameters from modules.call_queue import wrap_gradio_gpu_call, wrap_queued_call, wrap_gradio_call import modules NAME = 'RestartLab' CFGDENOISER =["CFGDenoiserKDiffusion","CFGDenoiserLCM"] SAMPLING = ["Heun", "Euler","Euler a", "LCM","DPM++ SDE","DPM++ 2M","DPM++ 2M SDE","DPM++ 2M SDE Heun","DPM++ 3M SDE"] SAMPLING.extend(["Noise"]) NOISE =["SDE","Ancestral","Brownian","Brownian remake"] CFGMODE =["OFF","Constant","After","Linear","Slow2","Slow4","Slow6","Fast2","Fast4","Fast6","Turn2","Turn4","Turn6"] AUTOSKIP = ["Off","100%","80%","60%","50%","40%","Custom","Next","1st","2nd","3rd","4th","5th"] _restart_steps = 6 _restart_cfgDenoiser = "CFGDenoiserKDiffusion" _restart_sampling = "Heun" _restart_firstWeight = 1 _restart_noise = "SDE" _restart_changeCFG = 7.0 _restart_changeCFGstep = 0.0 _restart_changeCFGmode = "After" _restart_Start_s_churn = 0.0 _restart_End_s_churn = 0.0 _restart_s_churn_step = 0.0 _restart_s_churn_mode = "OFF" _restart_Start_s_noise = 0.0 _restart_End_s_noise = 0.0 _restart_s_noise_step = 0.0 _restart_s_noise_mode = "OFF" _skipStep = 0 _myskip = False _usemetaFirst = False _savemeta = True _skipmeta = "" def round(a,x) : return int(a * 10 ** x ) / 10 ** x def lerp(a,b,t): return round(a + t * (b - a),3) def lerp2(x,step_id,startCFG,restart_changeCFG,restart_changeCFGmode,restart_changeCFGstep) : if restart_changeCFGmode == "After" : if step_id == restart_changeCFGstep : x = restart_changeCFG elif step_id < restart_changeCFGstep and restart_changeCFGstep > 1 : if restart_changeCFGmode == "Linear": x = lerp(startCFG,restart_changeCFG, (step_id / (restart_changeCFGstep - 1))) elif restart_changeCFGmode == "Slow2": x = lerp(startCFG,restart_changeCFG, (step_id / (restart_changeCFGstep - 1)) ** 2.0 ) elif restart_changeCFGmode == "Slow4": x = lerp(startCFG,restart_changeCFG, (step_id / (restart_changeCFGstep - 1)) ** 4.0 ) elif restart_changeCFGmode == "Slow6": x = lerp(startCFG,restart_changeCFG, (step_id / (restart_changeCFGstep - 1)) ** 6.0 ) elif restart_changeCFGmode == "Fast2": x = round(restart_changeCFG + (startCFG - restart_changeCFG) * ((1 - step_id / (restart_changeCFGstep - 1) ) ** 2.0),2) elif restart_changeCFGmode == "Fast4": x = round(restart_changeCFG + (startCFG - restart_changeCFG) * ((1 - step_id / (restart_changeCFGstep - 1) ) ** 4.0),2) elif restart_changeCFGmode == "Fast6": x = round(restart_changeCFG + (startCFG - restart_changeCFG) * ((1 - step_id / (restart_changeCFGstep - 1) ) ** 6.0),2) if restart_changeCFGmode == "Turn2": val = round(restart_changeCFG + (startCFG - restart_changeCFG) * ((1 - step_id / (restart_changeCFGstep - 1) ) ** 2.0),2) x = val if min(startCFG , restart_changeCFG) <= val and val <= max(startCFG , restart_changeCFG) else startCFG elif restart_changeCFGmode == "Turn4": val = round(restart_changeCFG + (startCFG - restart_changeCFG) * ((1 - step_id / (restart_changeCFGstep - 1) ) ** 4.0),2) x = val if min(startCFG , restart_changeCFG) <= val and val <= max(startCFG , restart_changeCFG) else startCFG elif restart_changeCFGmode == "Turn6": val = round(restart_changeCFG + (startCFG - restart_changeCFG) * ((1 - step_id / (restart_changeCFGstep - 1) ) ** 6.0),2) x = val if min(startCFG , restart_changeCFG) <= val and val <= max(startCFG , restart_changeCFG) else startCFG return x @torch.no_grad() def restartLab_sampler(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2, s_churn=0., s_tmin=0., s_tmax=float('inf'), restart_list=None): """Implements restart sampling in Restart Sampling for Improving Generative Processes (2023) Restart_list format: {min_sigma: [ restart_steps, restart_times, max_sigma]} If restart_list is None: will choose restart_list automatically, otherwise will use the given restart_list """ global _restart_steps,_restart_sampling,_restart_cfgDenoiser,_restart_firstWeight,_restart_noise,_restart_changeCFG,_restart_changeCFGstep,_restart_changeCFGmode,_restart_Start_s_churn,_restart_End_s_churn,_restart_s_churn_step,_restart_s_churn_mode,_skipStep,_skipmeta,_myskip,_restart_Start_s_noise ,_restart_End_s_noise ,_restart_s_noise_step ,_restart_s_noise_mode,_usemetaFirst restart_steps = _restart_steps restart_sampling = _restart_sampling restart_cfgDenoiser = _restart_cfgDenoiser restart_firstWeight = _restart_firstWeight restart_noise = _restart_noise restart_changeCFG = _restart_changeCFG restart_changeCFGstep = _restart_changeCFGstep restart_changeCFGmode = _restart_changeCFGmode restart_Start_s_churn = _restart_Start_s_churn restart_End_s_churn = _restart_End_s_churn restart_s_churn_step = _restart_s_churn_step restart_s_churn_mode = _restart_s_churn_mode restart_Start_s_noise = _restart_Start_s_noise restart_End_s_noise = _restart_End_s_noise restart_s_noise_step = _restart_s_noise_step restart_s_noise_mode = _restart_s_noise_mode if restart_changeCFGstep <= 1 : restart_changeCFGstep = int((len(sigmas) - 1) * restart_changeCFGstep) if restart_s_churn_step <= 1 : restart_s_churn_step = int((len(sigmas) - 1) * restart_s_churn_step) if restart_s_noise_step <= 1 : restart_s_noise_step = int((len(sigmas) - 1) * restart_s_noise_step) if _skipStep <= 1 and _skipStep > 0 : _skipStep = int((len(sigmas) - 1) * _skipStep) if restart_changeCFGstep > 1 : restart_changeCFGstep = int(restart_changeCFGstep) if restart_s_churn_step > 1 : restart_s_churn_step = int(restart_s_churn_step) if restart_s_noise_step > 1 : restart_s_noise_step = int(restart_s_noise_step) if _skipStep > 1 : _skipStep = int(_skipStep) extra_args = {} if extra_args is None else extra_args s_in = x.new_ones([x.shape[0]]) step_id = 0 from k_diffusion.sampling import to_d, get_sigmas_karras,get_ancestral_step ################################# # sampling def heun_step(x, old_sigma, new_sigma, second_order=True): nonlocal step_id denoised = model(x, old_sigma * s_in, **extra_args) d = to_d(x, old_sigma, denoised) if callback is not None: callback({'x': x, 'i': step_id, 'sigma': new_sigma, 'sigma_hat': old_sigma, 'denoised': denoised}) dt = new_sigma - old_sigma if new_sigma == 0 or not second_order: # Euler method x = x + d * dt else: # Heun's method x_2 = x + d * dt denoised_2 = model(x_2, new_sigma * s_in, **extra_args) d_2 = to_d(x_2, new_sigma, denoised_2) d_prime = (d + d_2) / 2 x = x + d_prime * dt step_id += 1 return x def euler_step(x, old_sigma,new_sigma,sgima_size): nonlocal step_id """Implements Algorithm 2 (Euler steps) from Karras et al. (2022).""" denoised = model(x, old_sigma * s_in, **extra_args) d = to_d(x, old_sigma, denoised) if callback is not None: callback({'x': x, 'i': step_id, 'sigma': new_sigma, 'sigma_hat': old_sigma, 'denoised': denoised}) dt = new_sigma - old_sigma # Euler method x = x + d * dt step_id += 1 return x def euler_ancestral_step( x, old_sigma,new_sigma, eta=1., s_noise=1., noise_sampler=None): """Ancestral sampling with Euler method steps.""" nonlocal step_id #extra_args = {} if extra_args is None else extra_args #noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler #s_in = x.new_ones([x.shape[0]]) denoised = model(x, old_sigma * s_in, **extra_args) sigma_down, sigma_up = get_ancestral_step(old_sigma,new_sigma, eta=eta) if callback is not None: callback({'x': x, 'i': step_id, 'sigma': new_sigma, 'sigma_hat': old_sigma, 'denoised': denoised}) d = to_d(x, old_sigma, denoised) # Euler method dt = sigma_down - old_sigma x = x + d * dt if new_sigma > 0: x = x + noise_sampler(old_sigma, new_sigma) * s_noise * sigma_up step_id += 1 return x def lcm_step(x,old_sigma,new_sigma,noise_sampler=None): nonlocal step_id #noise_sampler = default_noise_sampler(x) denoised = model(x, old_sigma * s_in, **extra_args) if callback is not None: callback({'x': x, 'i': step_id, 'sigma': new_sigma, 'sigma_hat': old_sigma, 'denoised': denoised}) x = denoised if new_sigma > 0: x += new_sigma * noise_sampler(old_sigma, new_sigma) step_id += 1 return x def dpmpp_sde_step(x, sigmas, eta=1., s_noise=1., noise_sampler=None, r=1 / 2): """DPM-Solver++ (stochastic).""" nonlocal step_id sigma_fn = lambda t: t.neg().exp() t_fn = lambda sigma: sigma.log().neg() i = 0 denoised = model(x, sigmas[i] * s_in, **extra_args) if callback is not None: callback({'x': x, 'i': step_id, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) if sigmas[i + 1] == 0: # Euler method d = to_d(x, sigmas[i], denoised) dt = sigmas[i + 1] - sigmas[i] x = x + d * dt else: # DPM-Solver++ t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1]) h = t_next - t s = t + h * r fac = 1 / (2 * r) # Step 1 sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(s), eta) s_ = t_fn(sd) x_2 = (sigma_fn(s_) / sigma_fn(t)) * x - (t - s_).expm1() * denoised x_2 = x_2 + noise_sampler(sigma_fn(t), sigma_fn(s)) * s_noise * su denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args) # Step 2 sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(t_next), eta) t_next_ = t_fn(sd) denoised_d = (1 - fac) * denoised + fac * denoised_2 x = (sigma_fn(t_next_) / sigma_fn(t)) * x - (t - t_next_).expm1() * denoised_d x = x + noise_sampler(sigma_fn(t), sigma_fn(t_next)) * s_noise * su step_id += 1 return x def dpmpp_2m_step(x,sigmas,last_sigma): """DPM-Solver++(2M).""" nonlocal step_id ,old_denoised sigma_fn = lambda t: t.neg().exp() t_fn = lambda sigma: sigma.log().neg() i = 0 denoised = model(x, sigmas[i] * s_in, **extra_args) if callback is not None: callback({'x': x, 'i': step_id, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1]) h = t_next - t if old_denoised is None or sigmas[i + 1] == 0: x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised else: h_last = t - t_fn(last_sigma) #t_fn(sigmas[i - 1]) r = h_last / h denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_d old_denoised = denoised step_id += 1 return x def dpmpp_2m_test_step(x,sigmas,last_sigma): """DPM-Solver++(2M).""" # cf. https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8457 nonlocal step_id ,old_denoised sigma_fn = lambda t: t.neg().exp() t_fn = lambda sigma: sigma.log().neg() i = 0 denoised = model(x, sigmas[i] * s_in, **extra_args) if callback is not None: callback({'x': x, 'i': step_id, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1]) h = t_next - t t_min = min(sigma_fn(t_next), sigma_fn(t)) t_max = max(sigma_fn(t_next), sigma_fn(t)) if old_denoised is None or sigmas[i + 1] == 0: x = (t_min / t_max) * x - (-h).expm1() * denoised else: h_last = t - t_fn(last_sigma) h_min = min(h_last, h) h_max = max(h_last, h) r = h_max / h_min h_d = (h_max + h_min) / 2 denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised x = (t_min / t_max) * x - (-h_d).expm1() * denoised_d old_denoised = denoised step_id += 1 return x def dpmpp_2m_test_v2n_step(x,sigmas,last_sigma,steps): """DPM-Solver++(2M).""" # cf. https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8457 nonlocal step_id ,old_denoised sigma_fn = lambda t: t.neg().exp() t_fn = lambda sigma: sigma.log().neg() i = 0 denoised = model(x, sigmas[i] * s_in, **extra_args) if callback is not None: callback({'x': x, 'i': step_id, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1]) h = t_next - t t_min = min(sigma_fn(t_next), sigma_fn(t)) t_max = max(sigma_fn(t_next), sigma_fn(t)) if old_denoised is None or sigmas[i + 1] == 0: x = (t_min / t_max) * x - (-h).expm1() * denoised else: h_last = t - t_fn(last_sigma) h_min = min(h_last, h) h_max = max(h_last, h) r = h_max / h_min h_d = h #(h_max + h_min) / 2 denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised x = (t_min / t_max) * x - (-h_d).expm1() * denoised_d #old_denoised = denoised #test new sampler sigma_progress = step_id / steps #len(sigmas) #doto steps adjustment_factor = 1 - (0.11 * (sigma_progress * sigma_progress)) old_denoised = denoised * adjustment_factor step_id += 1 return x def dpmpp_2m_sde_step(x, sigmas, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'): """DPM-Solver++(2M) SDE.""" nonlocal step_id ,old_denoised,h_last n = None i = 0 if solver_type not in {'heun', 'midpoint'}: raise ValueError('solver_type must be \'heun\' or \'midpoint\'') denoised = model(x, sigmas[i] * s_in, **extra_args) if callback is not None: callback({'x': x, 'i': step_id, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) if sigmas[i + 1] == 0: # Denoising step x = denoised t, s = -sigmas[i].log(), -sigmas[i + 1].log() h = s - t else: # DPM-Solver++(2M) SDE t, s = -sigmas[i].log(), -sigmas[i + 1].log() h = s - t eta_h = eta * h x = sigmas[i + 1] / sigmas[i] * (-eta_h).exp() * x + (-h - eta_h).expm1().neg() * denoised if old_denoised is not None: r = h_last / h if solver_type == 'heun': x = x + ((-h - eta_h).expm1().neg() / (-h - eta_h) + 1) * (1 / r) * (denoised - old_denoised) elif solver_type == 'midpoint': x = x + 0.5 * (-h - eta_h).expm1().neg() * (1 / r) * (denoised - old_denoised) x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * eta_h).expm1().neg().sqrt() * s_noise old_denoised = denoised h_last = h step_id += 1 return x denoised_1, denoised_2 = None, None h_1, h_2 = None, None def dpmpp_3m_sde_step(x, sigmas, eta=1., s_noise=1., noise_sampler=None): """DPM-Solver++(3M) SDE.""" nonlocal step_id ,denoised_1, denoised_2,h_1, h_2 i = 0 denoised = model(x, sigmas[i] * s_in, **extra_args) if callback is not None: callback({'x': x, 'i': step_id, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) if sigmas[i + 1] == 0: # Denoising step x = denoised t, s = -sigmas[i].log(), -sigmas[i + 1].log() h = s - t else: t, s = -sigmas[i].log(), -sigmas[i + 1].log() h = s - t h_eta = h * (eta + 1) x = torch.exp(-h_eta) * x + (-h_eta).expm1().neg() * denoised if h_2 is not None: r0 = h_1 / h r1 = h_2 / h d1_0 = (denoised - denoised_1) / r0 d1_1 = (denoised_1 - denoised_2) / r1 d1 = d1_0 + (d1_0 - d1_1) * r0 / (r0 + r1) d2 = (d1_0 - d1_1) / (r0 + r1) phi_2 = h_eta.neg().expm1() / h_eta + 1 phi_3 = phi_2 / h_eta - 0.5 x = x + phi_2 * d1 - phi_3 * d2 elif h_1 is not None: r = h_1 / h d = (denoised - denoised_1) / r phi_2 = h_eta.neg().expm1() / h_eta + 1 x = x + phi_2 * d if eta: x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * h * eta).expm1().neg().sqrt() * s_noise denoised_1, denoised_2 = denoised, denoised_1 h_1, h_2 = h, h_1 step_id += 1 return x #print(restart_steps) #print(restart_cfgDenoiser) #print(restart_sampling) #print(shared.opts.data.get('restartLab_steps',"None1") ) #print(shared.opts.data.get('restartLab_sampler',"None2") ) #print(shared.opts.data.get('restartLab_sampling',"None3") ) ################################# #AutoSkip startStep = 0 s = {} def base642numpy(data) : out_buf = zlib.decompress(base64.b64decode(data.encode())) return numpy.load(io.BytesIO(out_buf),allow_pickle=False) if not _skipmeta == "" : _skipmeta = _skipmeta.replace('^-^',',') s = ast.literal_eval(_skipmeta) startStep = s["skipStep"] if _usemetaFirst : startStep = 0 x = torch.from_numpy(base642numpy(s["skipMeta"])) _skipmeta = "" # restart steps = sigmas.shape[0] - 1 if restart_list is None: if steps >= restart_steps * 2: restart_times = steps // restart_steps - restart_firstWeight if restart_times <= 0 : restart_times = 1 sigmas = get_sigmas_karras(steps - restart_steps * restart_times, sigmas[-2].item(), sigmas[0].item(), device=sigmas.device) restart_list = {0.1: [restart_steps + 1, restart_times, 2]} else: restart_list = {} restart_list = {int(torch.argmin(abs(sigmas - key), dim=0)): value for key, value in restart_list.items()} firstCheck = False step_list = [] for i in range(len(sigmas) - 1): step_list.append((sigmas[i], sigmas[i + 1])) if not firstCheck : if i + 1 in restart_list: restart_steps, restart_times, restart_max = restart_list[i + 1] min_idx = i + 1 max_idx = int(torch.argmin(abs(sigmas - restart_max), dim=0)) i2 = i if max_idx < min_idx: sigma_restart = get_sigmas_karras(restart_steps, sigmas[min_idx].item(), sigmas[max_idx].item(), device=sigmas.device)[:-1] while restart_times > 0: if _skipStep == -999 : if startStep < i2 : _skipStep = i2 elif _skipStep <= -2 : _skipStep += 1 elif _skipStep == -1 : _skipStep = i2 i2 += len(sigma_restart)-1 restart_times -= 1 step_list.extend([(old_sigma, new_sigma) for (old_sigma, new_sigma) in zip(sigma_restart[:-1], sigma_restart[1:])]) last_sigma = None old_denoised = None h_last = None noise_x = None if restart_changeCFGstep > 0 : startCFG = extra_args['cond_scale'] if restart_noise == "Ancestral" : noise_sampler = default_noise_sampler(x) elif restart_noise == "Brownian" or restart_noise == "Brownian remake" or noise_sampler is None: sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max() noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max) sigma2 = [] for old_sigma, new_sigma in tqdm.tqdm(step_list, disable=disable): #CFG if restart_changeCFGstep > 0 : extra_args['cond_scale'] = lerp2(extra_args['cond_scale'] ,step_id,startCFG,restart_changeCFG,restart_changeCFGmode,restart_changeCFGstep) #s_churn if restart_s_churn_mode == "Constant" : s_churn = restart_Start_s_churn elif _restart_s_churn_step > 0 and not (restart_s_churn_mode == "OFF") and (restart_Start_s_churn > 0 or restart_End_s_churn > 0): s_churn = lerp2(s_churn,step_id,restart_Start_s_churn,restart_End_s_churn,restart_s_churn_mode,restart_s_churn_step) #s_noise if restart_s_noise_mode == "Constant" : s_noise = restart_Start_s_noise elif _restart_s_noise_step > 0 and not (restart_s_noise_mode == "OFF") and (restart_Start_s_noise > 0 or restart_End_s_noise > 0): s_noise = lerp2(s_noise,step_id,restart_Start_s_noise,restart_End_s_noise,restart_s_noise_mode,restart_s_noise_step) #Noise if last_sigma is None: last_sigma = old_sigma last_sigma2 = last_sigma elif last_sigma < old_sigma: x = x + k_diffusion.sampling.torch.randn_like(x) * s_noise * (old_sigma ** 2 - last_sigma ** 2) ** 0.5 if restart_noise == "Brownian remake" : if step_id < startStep and startStep > 0 : if 'noise_x' in s : x = torch.from_numpy(base642numpy(s["noise_x"])) if step_id < _skipStep : noise_x = x sigma_min, sigma_max = sigma_restart[sigma_restart > 0].min(), sigma_restart.max() noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max) #sigma churn gamma = min(s_churn / (len(step_list)), 2 ** 0.5 - 1) if s_tmin <= new_sigma <= s_tmax else 0. if gamma > 0 : eps = torch.randn_like(x) * s_noise sigma_hat = new_sigma * (gamma + 1) x = x + eps * (sigma_hat ** 2 - new_sigma ** 2) ** 0.5 #sampling if startStep > step_id: #AutoSkip if callback is not None: callback({'x': x, 'i': step_id, 'sigma': old_sigma, 'sigma_hat': new_sigma, 'denoised': 0}) step_id += 1 if step_id == startStep : x = torch.from_numpy(base642numpy(s["skipMeta"])) if 'old_denoised' in s : old_denoised = torch.from_numpy(base642numpy(s["old_denoised"])) if 'h_last' in s : h_last = torch.from_numpy(base642numpy(s["h_last"])) if 'denoised_1' in s : denoised_1 = torch.from_numpy(base642numpy(s["denoised_1"])) if 'denoised_2' in s : denoised_2 = torch.from_numpy(base642numpy(s["denoised_2"])) if 'h_1' in s : h_1 = torch.from_numpy(base642numpy(s["h_1"])) if 'h_2' in s : h_2 = torch.from_numpy(base642numpy(s["h_2"])) elif restart_sampling == "Euler": x = euler_step(x, old_sigma, new_sigma,len(step_list)) elif restart_sampling == "Euler a": x = euler_ancestral_step(x, old_sigma, new_sigma,eta,s_noise,noise_sampler) elif restart_sampling == "LCM": x = lcm_step(x, old_sigma, new_sigma,noise_sampler) elif restart_sampling == "DPM++ SDE": x = dpmpp_sde_step(x,[ old_sigma, new_sigma],eta,s_noise,noise_sampler) elif restart_sampling == "DPM++ 2M": x = dpmpp_2m_step(x,[ old_sigma, new_sigma],last_sigma2) elif restart_sampling == "DPM++ 2M test": x = dpmpp_2m_test_step(x,[ old_sigma, new_sigma],last_sigma2) elif restart_sampling == "DPM++ 2M test v2n": x = dpmpp_2m_test_v2n_step(x,[ old_sigma, new_sigma],last_sigma2,len(step_list)) elif restart_sampling == "DPM++ 2M B": x = dpmpp_2m_step(x,[ old_sigma, new_sigma],new_sigma) #Bug for LCM elif restart_sampling == "DPM++ 2M SDE": x = dpmpp_2m_sde_step(x, [old_sigma, new_sigma],eta,s_noise,noise_sampler) elif restart_sampling == "DPM++ 2M SDE Heun" : x = dpmpp_2m_sde_step(x, [old_sigma, new_sigma],eta,s_noise,noise_sampler,'heun') elif restart_sampling == "DPM++ 3M SDE": x = dpmpp_3m_sde_step(x, [old_sigma, new_sigma],eta,s_noise,noise_sampler) elif restart_sampling == "Noise": step_id += 1 else: x = heun_step(x, old_sigma, new_sigma) last_sigma = new_sigma last_sigma2 = old_sigma #AutoSkip if step_id == _skipStep and _skipStep > 0 : def numpy2base64(data) : out_buf = io.BytesIO() numpy.save(out_buf,data,allow_pickle=False) b64 = base64.b64encode(zlib.compress(out_buf.getvalue(),9)).decode("utf-8") return b64 s = {"skipStep" : _skipStep, "skipMeta" : numpy2base64(x.numpy())} if not noise_x is None: s['noise_x'] = numpy2base64(noise_x.numpy()) if not old_denoised is None: s['old_denoised'] = numpy2base64(old_denoised.numpy()) if not h_last is None: s['h_last'] = numpy2base64(h_last.numpy()) if not denoised_1 is None: s['denoised_1'] = numpy2base64(denoised_1.numpy()) if not denoised_2 is None: s['denoised_2'] = numpy2base64(denoised_2.numpy()) if not h_1 is None: s['h_1'] = numpy2base64(h_1.numpy()) if not h_2 is None: s['h_2'] = numpy2base64(h_2.numpy()) _skipmeta = str(s).replace(',','^-^') #_x2 = torch.from_numpy(base642numpy(d["skipMeta"])) if step_id < len(step_list): state.skip() _myskip = True return x ################################# # UI class Script(scripts.Script): def title(self): return NAME def show(self, is_img2img): return scripts.AlwaysVisible def ui(self, is_img2img): global _restart_steps,_restart_sampling,_restart_cfgDenoiser,_restart_firstWeight,_restart_noise def initial_seteps(): return int(shared.opts.data.get("restartlab_steps", _restart_steps)) def initial_cfgdenoiser(): return shared.opts.data.get("restartlab_cfgdenoiser", _restart_cfgDenoiser) def initial_sampling(): return shared.opts.data.get("restartlab_sampling", _restart_sampling) def initial_firstweight(): return shared.opts.data.get("restartlab_firstweight", _restart_firstWeight) def initial_noise(): return shared.opts.data.get("restartlab_noise", _restart_noise) strImg2img = "txt" if is_img2img : strImg2img = "img" def getFile(files) : if type(files) == list : return len(files) return -2 def loadImg(files): if files is None : return None,None,0 elif type(files) == list : if len(files) > 0: item = files.pop(0) items = [] for file in files: items.append(file.name) if len(items) > 0 : return item.name,items,len(items) else : return item.name,None,0 else : return files.name,None,0 return None,None,0 def infoFilter2 (geninfo,png_info_props): for item in png_info_props: if "RestartLab Sampler Parameters" == item : item = "RestartLab changeCFG" st = geninfo.find(item + ":") if st >= 0 : ed = geninfo.find(",",st) + 1 st2 = st if "RestartLab changeCFG" == item : st2 = geninfo.find("RestartLab s_noise_mode:",st) ed = geninfo.find(",",st2) + 1 geninfo = geninfo[:st] + geninfo[ed:] return geninfo with gr.Accordion(NAME, open=False): with gr.Column(): with gr.Row(): msl_firstWeight = gr.Slider(minimum=1, maximum=10, step=1, value=initial_firstweight, label='first Weight',min_width=120) msl_steps = gr.Slider(minimum=2, maximum=30, step=1, value=initial_seteps, label='restart Step',min_width=120) with gr.Row(): msl_cfgDenoiser = gr.Dropdown(CFGDENOISER, value=initial_cfgdenoiser, label="CFG Denoiser",min_width=120) msl_noise = gr.Dropdown(NOISE, value=initial_noise, label="Noise",min_width=120) msl_sampling = gr.Dropdown(SAMPLING, value=initial_sampling, label="Sampling",min_width=120) with gr.Accordion("Sampler Parameters", open=False): with gr.Column(): with gr.Row(): msl_changeCFG = gr.Slider(minimum=1.0, maximum=10.0, step=0.1, value=7.0, label='Change CFG',min_width=120) msl_ChangeCFGstep = gr.Slider(minimum=0.0, maximum=100.0, step=0.1, value=0.0, label='Change CFG Step',min_width=120) msl_ChangeCFGmode = gr.Dropdown(CFGMODE, value="After", label="Change CFG mode",min_width=120) with gr.Column(): with gr.Row(): msl_Start_s_churn = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, value=0.0, label='Start s_churn' ,min_width=120) msl_End_s_churn = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, value=0.0, label='End s_churn',min_width=120) msl_s_churn_step = gr.Slider(minimum=0.0, maximum=100.0, step=0.1, value=0.0, label='s_churn Step',min_width=120) msl_s_churn_mode = gr.Dropdown(CFGMODE, value="OFF", label="s_churn mode",min_width=120) with gr.Column(): with gr.Row(): msl_Start_s_noise = gr.Slider(minimum=0.0, maximum=2.0, step=0.001, value=1.0, label='Start s_noise' ,min_width=120) msl_End_s_noise = gr.Slider(minimum=0.0, maximum=2.0, step=0.001, value=1.0, label='End s_noise',min_width=120) msl_s_noise_step = gr.Slider(minimum=0.0, maximum=100.0, step=0.1, value=0.0, label='s_noise Step',min_width=120) msl_s_noise_mode = gr.Dropdown(CFGMODE, value="OFF", label="s_noise mode",min_width=120) with gr.Accordion("AutoSkip", open=False): skipmode = gr.Radio(AUTOSKIP, value="Off", label="Skip Mode") skipstep = gr.Slider(minimum=-10, maximum=30, step=0.1, value=0, label='Custom Step') with gr.Column(): with gr.Row(): savemeta = gr.Checkbox(value=True,label="Save SkipMeta to pnginfo") usemeta = gr.Checkbox(value=True, label='Use SkipMeta') usemetaFirst = gr.Checkbox(value=False, label='Make the first step SkipMeta') skipmeta = gr.Text(value="", label='SkipMeta' ,elem_id=strImg2img + "skipMeta",max_lines=3) with gr.Accordion("Batch", open=False): with gr.Column(): with gr.Row(): startBtn = gr.Button(value="Start" ,elem_id=strImg2img + "BatchStartButton",variant="primary") stopBtn = gr.Button(value="Stop" ,elem_id=strImg2img + "BatchStopButton",variant="secondary") with gr.Row(): img2img_batch_png_info_props = gr.CheckboxGroup(["Steps", "Denoising strength","RestartLab firstweight","RestartLab steps", "RestartLab noise","RestartLab sampling","RestartLab Sampler Parameters"], label="Parameters to remove from png info") files = gr.File(label="Batch Process", file_count="multiple", interactive=True, type="file", elem_id=self.elem_id("files")) nextfile = gr.Number(label="Files", visible=True, elem_id=strImg2img + "nextFile") with gr.Accordion("Tools", open=False): with gr.Column(): with gr.Row(): btn = gr.Button(visible=True,value="Next" ,elem_id=strImg2img + "nextButton") listBtn = gr.Button(value="List" ,elem_id=strImg2img + "BatchListButton") btnPnginfo = gr.Button(visible=True,value="Pnginfo" ,elem_id=strImg2img + "pnginfoButton") buttons = parameters_copypaste.create_buttons([strImg2img +"2img"]) image = gr.Image(visible=True,elem_id="pnginfo_image2", label="Source", source="upload", interactive=True, type="pil") html = gr.Textbox(visible=False, elem_id="pnginfo_generation_info1") generation_info = gr.Textbox(visible=True,label="PngInfo" ,max_lines=5, elem_id=strImg2img + "pnginfo_generation_info3") html2 = gr.Textbox(visible=False, elem_id="pnginfo_generation_info2") btn.click(loadImg, inputs=[files], outputs=[image,files,nextfile]) for tabname, button in buttons.items(): #button.visible = False parameters_copypaste.register_paste_params_button(parameters_copypaste.ParamBinding( paste_button=button, tabname=tabname, source_text_component=generation_info, source_image_component=image, )) def run_pnginfo2(image,png_info_props) : t , geninfo ,info = modules.extras.run_pnginfo(image) geninfo = infoFilter2(geninfo,png_info_props) return '' , geninfo image.change( fn=wrap_gradio_call(run_pnginfo2), inputs=[image,img2img_batch_png_info_props], outputs=[html,generation_info], ) btnPnginfo.click( #fn=wrap_gradio_call(modules.extras.run_pnginfo), fn=wrap_gradio_call(run_pnginfo2), inputs=[image,img2img_batch_png_info_props], outputs=[html,generation_info], ) def getFilelist(files): if files is None : return 0 elif type(files) == list : return len(files) else : return 0 listBtn.click( fn=getFilelist, inputs=[files], outputs=[nextfile], ) files.change( fn=getFilelist, inputs=[files], outputs=[nextfile], ) strHtml = """ function() { var lastCount = -1, lastSeed = -1, loopPhase = 0, strImg = '""" + strImg2img + """'; let genbutton = gradioApp().querySelector("#" + strImg + "2img_generate"); console.log(genbutton); let interruptbutton = gradioApp().querySelector("#" + strImg + "2img_interrupt"); console.log(interruptbutton); let nextButton = gradioApp().querySelector("#" + strImg + "nextButton"); console.log(nextButton); let stopButton = gradioApp().querySelector("#" + strImg + "BatchStopButton"); console.log(stopButton); let startButton = gradioApp().querySelector("#" + strImg + "BatchStartButton"); startButton.classList.add('secondary'); stopButton.classList.add('primary'); startButton.classList.remove('primary'); stopButton.classList.remove('secondary'); let listButton = gradioApp().querySelector("#" + strImg + "BatchListButton"); console.log(listButton); let nextFile = gradioApp().querySelector("#" + strImg + "nextFile input"); console.log(nextFile); let pngInfoBtn = gradioApp().querySelector("#" + strImg + "pnginfoButton"); console.log(pngInfoBtn); let pngInfo = gradioApp().querySelector("#" + strImg + "pnginfo_generation_info3 textarea"); console.log(pngInfo); let txt2button = gradioApp().querySelector("#" + strImg + "2img_tab"); console.log(txt2button); let seedInput = gradioApp().querySelector("#" + strImg + "2img_seed input"); let skipMeta = gradioApp().querySelector("#" + strImg + "skipMeta textarea"); console.log(skipMeta); console.log(interruptbutton.offsetParent); interruptbutton.offsetParent ? loopPhase = 3 : ( lastCount = -2, loopPhase = 0); clearInterval(window.generateOnRepeatInterval2); window.generateOnRepeatInterval2 = setInterval(function() { 0 == loopPhase ? ( console.log("rest"), console.log(lastCount), 0 < lastCount || lastCount == -2 ? ( skipMeta.value = "", skipMeta.dispatchEvent(new Event("input", {bubbles: !0})), seedInput.value = "-2", seedInput.dispatchEvent(new Event("input", {bubbles: !0})), nextFile.value = -2, nextFile.dispatchEvent(new Event("input", {bubbles: !0})), pngInfo.value = "", pngInfo.dispatchEvent(new Event("input", {bubbles: !0})), loopPhase = 1 ) : ( console.log("Exit"), clearInterval(window.generateOnRepeatInterval2), stopButton.click() ) ) : 1 == loopPhase ? ( console.log("next"), console.log(nextFile.value), console.log(seedInput.value), -2 == seedInput.value && -2 == nextFile.value && "" == skipMeta.value && "" == pngInfo.value && ( nextButton.click(), loopPhase = 2 ) ) : 2 == loopPhase ? ( console.log("info"), nextFile.value != lastCount && nextFile.value != -2 ? ( lastCount = nextFile.value, console.log(lastCount), loopPhase = 3 ) : listButton.click() ) : 3 == loopPhase ? ( console.log("pnginfo"), console.log(seedInput.value), pngInfo.value != "" ? ( txt2button.click(), loopPhase = 4 ) : pngInfoBtn.click() ) : 6 >= loopPhase ? ( console.log("seed"), console.log(seedInput.value), console.log(loopPhase), seedInput.value != lastSeed && seedInput.value != -2 ? ( loopPhase = 7 ) : ( seedInput.value == lastSeed & (pngInfoBtn.click(), loopPhase += 1), txt2button.click() ) ) : 12 > loopPhase ? ( "" == skipMeta.value ? ( seedInput.value == lastSeed ? pngInfoBtn.click() : txt2button.click(), loopPhase += 1 ) : ( loopPhase = 12, lastSeed = seedInput.value, console.log(lastSeed) ), console.log("wait" + loopPhase) ) : 12 == loopPhase ? ( console.log("generate"), interruptbutton.style.display == "block" ? loopPhase = 13 : genbutton.click(), console.log(interruptbutton.style.display) ) : 13 != loopPhase || interruptbutton.style.display == "block" || ( loopPhase = 0, console.log(interruptbutton.style.display) ) }, 2000); } """ strHtml = strHtml.replace(" ","") strHtml = strHtml.replace("\t","") strHtml = strHtml.replace('\n', '') startBtn.click( fn=None, _js=strHtml, inputs=None, outputs=None, ) stopBtn.click( fn=None, _js="function(){strImg = '" + strImg2img + "';clearInterval(window.generateOnRepeatInterval2);let stopButton = gradioApp().querySelector('#' + strImg + 'BatchStopButton');let startButton = gradioApp().querySelector('#' + strImg + 'BatchStartButton');stopButton.classList.add('secondary');startButton.classList.add('primary');stopButton.classList.remove('primary');startButton.classList.remove('secondary');}", inputs=None, outputs=None, ) self._img2img = is_img2img self.infotext_fields = ( (msl_steps, "RestartLab steps"), (msl_firstWeight, "RestartLab firstweight"), (msl_cfgDenoiser, "RestartLab cfgDenoiser"), (msl_sampling, "RestartLab sampling"), (msl_noise, "RestartLab noise"), (msl_changeCFG, "RestartLab changeCFG"), (msl_ChangeCFGstep, "RestartLab changeCFGstep"), (msl_ChangeCFGmode, "RestartLab changeCFGmode"), (msl_Start_s_churn, "RestartLab stat_s_churn"), (msl_End_s_churn, "RestartLab end_s_churn"), (msl_s_churn_step, "RestartLab s_churn_step"), (msl_s_churn_mode, "RestartLab s_churn_mode"), (msl_Start_s_noise, "RestartLab stat_s_noise"), (msl_End_s_noise, "RestartLab end_s_noise"), (msl_s_noise_step, "RestartLab s_noise_step"), (msl_s_noise_mode, "RestartLab s_noise_mode"), (skipmeta, "AutoSkipMeta"), ) return [msl_steps,msl_cfgDenoiser,msl_sampling,msl_firstWeight,msl_noise,msl_changeCFG,msl_ChangeCFGstep,msl_ChangeCFGmode,msl_Start_s_churn,msl_End_s_churn,msl_s_churn_step,msl_s_churn_mode,msl_Start_s_noise,msl_End_s_noise,msl_s_noise_step,msl_s_noise_mode,skipmode,skipstep,savemeta,usemeta,usemetaFirst,skipmeta] def process(self, p,msl_steps,msl_cfgDenoiser,msl_sampling,msl_firstWeight,msl_noise,msl_changeCFG,msl_ChangeCFGstep,msl_ChangeCFGmode,msl_Start_s_churn,msl_End_s_churn,msl_s_churn_step,msl_s_churn_mode,msl_Start_s_noise,msl_End_s_noise,msl_s_noise_step,msl_s_noise_mode,skipmode,skipstep,savemeta,usemeta,usemetaFirst,skipmeta): global _restart_steps,_restart_sampling,_restart_cfgDenoiser,_restart_firstWeight,_restart_noise,_restart_changeCFG,_restart_changeCFGstep,_restart_changeCFGmode,_restart_Start_s_churn,_restart_End_s_churn,_restart_s_churn_step,_restart_s_churn_mode,_restart_Start_s_noise ,_restart_End_s_noise ,_restart_s_noise_step ,_restart_s_noise_mode,_skipStep,_skipmeta,_usemetaFirst,_savemeta if p.sampler_name == NAME or p.sampler_name == NAME + " karras" or p.sampler_name == NAME + " Exponential": #xyz _restart_steps = getattr(p, NAME + "_steps", msl_steps) _restart_firstWeight = getattr(p, NAME + "_firstWeight", msl_firstWeight) _restart_cfgDenoiser = getattr(p, NAME + "_cfgDenoiser", msl_cfgDenoiser) _restart_sampling = getattr(p, NAME + "_sampling", msl_sampling) _restart_noise = getattr(p, NAME + "_noise", msl_noise) _restart_changeCFG = getattr(p, NAME + "_changeCFG", msl_changeCFG) _restart_changeCFGstep = getattr(p, NAME + "_changeCFGstep", msl_ChangeCFGstep) _restart_changeCFGmode = getattr(p, NAME + "_changeCFGmode", msl_ChangeCFGmode) _restart_Start_s_churn = getattr(p, NAME + "_stat_s_churn", msl_Start_s_churn) _restart_End_s_churn = getattr(p, NAME + "_end_s_churn", msl_End_s_churn) _restart_s_churn_step = getattr(p, NAME + "_s_churn_step", msl_s_churn_step) _restart_s_churn_mode = getattr(p, NAME + "_s_churn_mode", msl_s_churn_mode) _restart_Start_s_noise = getattr(p, NAME + "_stat_s_noise", msl_Start_s_noise) _restart_End_s_noise = getattr(p, NAME + "_end_s_noise", msl_End_s_noise) _restart_s_noise_step = getattr(p, NAME + "_s_noise_step", msl_s_noise_step) _restart_s_noise_mode = getattr(p, NAME + "_s_noise_mode", msl_s_noise_mode) self._skipmode = getattr(p, 'autoskip_skipmode', skipmode) _skipStep = getattr(p, NAME + "_AutoSkip_Step", skipstep) _skipmeta = skipmeta _usemetaFirst = usemetaFirst if _skipStep == 1 : _skipStep += 0.1 if(self._skipmode == 'Off'): _skipStep = 0 elif(self._skipmode == '1st'): _skipStep = -1 elif(self._skipmode == '2nd'): _skipStep = -2 elif(self._skipmode == '3rd'): _skipStep = -3 elif(self._skipmode == '4th'): _skipStep = -4 elif(self._skipmode == '5th'): _skipStep = -5 elif(self._skipmode == 'Next'): _skipStep = -999 elif(self._skipmode == '100%'): _skipStep = 1.0 elif(self._skipmode == '80%'): _skipStep = 0.8 elif(self._skipmode == '60%'): _skipStep = 0.6 elif(self._skipmode == '50%'): _skipStep = 0.5 elif(self._skipmode == '40%'): _skipStep = 0.4 if not usemeta : _skipmeta = "" _savemeta = savemeta #bug fix if "Branian" in _restart_noise: _restart_noise = _restart_noise.replace('Branian', 'Brownian') p.extra_generation_params["RestartLab steps"] = _restart_steps p.extra_generation_params["RestartLab firstweight"] = _restart_firstWeight p.extra_generation_params["RestartLab cfgDenoiser"] = _restart_cfgDenoiser p.extra_generation_params["RestartLab sampling"] = _restart_sampling p.extra_generation_params["RestartLab noise"] = _restart_noise p.extra_generation_params["RestartLab changeCFG"] = _restart_changeCFG p.extra_generation_params["RestartLab changeCFGstep"] = _restart_changeCFGstep p.extra_generation_params["RestartLab changeCFGmode"] = _restart_changeCFGmode p.extra_generation_params["RestartLab stat_s_churn"] = _restart_Start_s_churn p.extra_generation_params["RestartLab end_s_churn"] = _restart_End_s_churn p.extra_generation_params["RestartLab s_churn_step"] = _restart_s_churn_step p.extra_generation_params["RestartLab s_churn_mode"] = _restart_s_churn_mode p.extra_generation_params["RestartLab stat_s_noise"] = _restart_Start_s_noise p.extra_generation_params["RestartLab end_s_noise"] = _restart_End_s_noise p.extra_generation_params["RestartLab s_noise_step"] = _restart_s_noise_step p.extra_generation_params["RestartLab s_noise_mode"] = _restart_s_noise_mode #shared.opts.data.update({'restartlab_steps':_restart_steps}) #shared.opts.data.update({'restartlab_cfgdenoiser':_restart_cfgDenoiser}) #shared.opts.data.update({'restartlab_sampling':_restart_sampling}) else: if("RestartLab steps" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab steps"] if("RestartLab firstweight" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab firstweight"] if("RestartLab cfgDenoiser" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab cfgDenoiser"] if("RestartLab sampling" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab sampling"] if("RestartLab noise" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab noise"] if("RestartLab changeCFG" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab changeCFG"] if("RestartLab changeCFGstep" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab changeCFGstep"] if("RestartLab changeCFGmode" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab changeCFGmode"] if("RestartLab stat_s_churn" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab stat_s_churn"] if("RestartLab end_s_churn" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab end_s_churn"] if("RestartLab s_churn_step" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab s_churn_step"] if("RestartLab s_churn_mode" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab s_churn_mode"] if("RestartLab stat_s_noise" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab stat_s_noise"] if("RestartLab end_s_noise" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab end_s_noise"] if("RestartLab s_noise_step" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab s_noise_step"] if("RestartLab s_noise_mode" in p.extra_generation_params.keys()): del p.extra_generation_params["RestartLab s_noise_mode"] on_before_image_saved(self.image_saved_callback) def process_batch(self, p, *args, **kwargs): global _myskip _myskip = False def postprocess_batch(self, p, *args, **kwargs): global _myskip if(_myskip): state.skipped = False def image_saved_callback(self,params): if 'parameters' not in params.pnginfo: return global _skipmeta,_savemeta if (not _skipmeta == ""): if _savemeta : strinfo = ", AutoSkipMeta: " + str(_skipmeta) params.pnginfo['parameters'] += strinfo _skipmeta = "" ################################# # xyz support def make_axis_on_xyz_grid(): xyz_grid = [x for x in scripts.scripts_data if x.script_class.__module__ == "xyz_grid.py"][0].module extra_axis_options = [ xyz_grid.AxisOption("["+NAME+"] steps", int,xyz_grid.apply_field(NAME + "_steps")), xyz_grid.AxisOption("["+NAME+"] firstWeight", int,xyz_grid.apply_field(NAME + "_firstWeight")), xyz_grid.AxisOption("["+NAME+"] cfgDenoiser", str, xyz_grid.apply_field(NAME + "_cfgDenoiser"), choices=lambda:CFGDENOISER), xyz_grid.AxisOption("["+NAME+"] sampling", str, xyz_grid.apply_field(NAME + "_sampling"), choices=lambda:SAMPLING), xyz_grid.AxisOption("["+NAME+"] noise", str, xyz_grid.apply_field(NAME + "_noise"), choices=lambda:NOISE), xyz_grid.AxisOption("["+NAME+"] changeCFG", float, xyz_grid.apply_field(NAME + "_changeCFG")), xyz_grid.AxisOption("["+NAME+"] changeCFGstep", float, xyz_grid.apply_field(NAME + "_changeCFGstep")), xyz_grid.AxisOption("["+NAME+"] changeCFGmode", str, xyz_grid.apply_field(NAME + "_changeCFGmode"), choices=lambda:CFGMODE), xyz_grid.AxisOption("["+NAME+"] stat s_churn", float, xyz_grid.apply_field(NAME + "_stat_s_churn")), xyz_grid.AxisOption("["+NAME+"] end s_churn", float, xyz_grid.apply_field(NAME + "_end_s_churn")), xyz_grid.AxisOption("["+NAME+"] s_churn step", float, xyz_grid.apply_field(NAME + "_s_churn_step")), xyz_grid.AxisOption("["+NAME+"] s_churn mode", str, xyz_grid.apply_field(NAME + "_s_churn_mode"), choices=lambda:CFGMODE), xyz_grid.AxisOption("["+NAME+"] stat s_noise", float, xyz_grid.apply_field(NAME + "_stat_s_noise")), xyz_grid.AxisOption("["+NAME+"] end s_noise", float, xyz_grid.apply_field(NAME + "_end_s_noise")), xyz_grid.AxisOption("["+NAME+"] s_noise step", float, xyz_grid.apply_field(NAME + "_s_noise_step")), xyz_grid.AxisOption("["+NAME+"] s_noise mode", str, xyz_grid.apply_field(NAME + "_s_noise_mode"), choices=lambda:CFGMODE), xyz_grid.AxisOption("["+NAME+"] AutoSkip CustomStep", float, xyz_grid.apply_field(NAME + "_AutoSkip_Step")), ] if not any("["+NAME+"]" in x.label for x in xyz_grid.axis_options): xyz_grid.axis_options.extend(extra_axis_options) def on_before_ui(): try: make_axis_on_xyz_grid() except Exception: error = traceback.format_exc() print( f" - " + NAME + " : xyz_grid error:\n{error}", file=sys.stderr, ) script_callbacks.on_before_ui(on_before_ui) ################################# # Setting def on_ui_settings(): section = ('RestartLab', "RestartLab") shared.opts.add_option( key="restartlab_steps", info=shared.OptionInfo( 6, "RestartLab steps", gr.Number, {"interactive": True}, section=section ), ) shared.opts.add_option( key="restartlab_firstweight", info=shared.OptionInfo( 1, "RestartLab firstweight", gr.Number, {"interactive": True}, section=section ), ) shared.opts.add_option( key="restartlab_cfgdenoiser", info=shared.OptionInfo( "CFGDenoiserKDiffusion", "RestartLab cfgDenoiser", gr.Radio, {"choices": CFGDENOISER}, section=section ), ) shared.opts.add_option( key="restartlab_sampling", info=shared.OptionInfo( "Heun", "RestartLab sampling", gr.Radio, {"choices": SAMPLING}, section=section ), ) shared.opts.add_option( key="restartlab_moise", info=shared.OptionInfo( "SDE", "RestartLab noise", gr.Radio, {"choices": NOISE}, section=section ), ) script_callbacks.on_ui_settings(on_ui_settings) ################################# # Sampler CFGDenoiser #if not NAME in [x.name for x in sd_samplers.all_samplers]: def add_RestartLab(): restartLab_samplers = [ (NAME, restartLab_sampler, ['k-restartLab'], {"second_order": True,"brownian_noise": True, }), (NAME + " karras", restartLab_sampler, ['k-restartLab'], {'scheduler': 'karras', "second_order": True,"brownian_noise": True, }), (NAME + " Exponential", restartLab_sampler, ['k-restartLab'], {'scheduler': 'exponential', "second_order": True,"brownian_noise": True, }), ] samplers_data_restartLab = [ sd_samplers_common.SamplerData(label, lambda model, funcname=funcname: restartLabSampler(funcname, model), aliases, options) for label, funcname, aliases, options in restartLab_samplers ] sd_samplers.all_samplers += samplers_data_restartLab sd_samplers.all_samplers_map = {x.name: x for x in sd_samplers.all_samplers} sd_samplers_kdiffusion.sampler_extra_params.setdefault(restartLab_sampler , ['s_churn', 's_tmin', 's_tmax', 's_noise']) sd_samplers.set_samplers() add_RestartLab() class restartLabSampler(sd_samplers_kdiffusion.KDiffusionSampler): def __init__(self, funcname, sd_model, options=None): super().__init__(funcname, sd_model, options) global _restart_cfgDenoiser if _restart_cfgDenoiser == "CFGDenoiserLCM": self.model_wrap_cfg = CFGDenoiserLCM(self) self.model_wrap = self.model_wrap_cfg.inner_model class LCMCompVisDenoiser(DiscreteEpsDDPMDenoiser): def __init__(self, model): timesteps = 1000 beta_start = 0.00085 beta_end = 0.012 betas = torch.linspace(beta_start**0.5, beta_end**0.5, timesteps, dtype=torch.float32) ** 2 alphas = 1.0 - betas alphas_cumprod = torch.cumprod(alphas, dim=0) original_timesteps = 50 # LCM Original Timesteps (default=50, for current version of LCM) self.skip_steps = timesteps // original_timesteps alphas_cumprod_valid = torch.zeros((original_timesteps), dtype=torch.float32, device=model.device) for x in range(original_timesteps): alphas_cumprod_valid[original_timesteps - 1 - x] = alphas_cumprod[timesteps - 1 - x * self.skip_steps] super().__init__(model, alphas_cumprod_valid, quantize=None) def get_sigmas(self, n=None, sgm=False): if n is None: return sampling.append_zero(self.sigmas.flip(0)) start = self.sigma_to_t(self.sigma_max) end = self.sigma_to_t(self.sigma_min) if sgm: t = torch.linspace(start, end, n + 1, device=shared.sd_model.device)[:-1] else: t = torch.linspace(start, end, n, device=shared.sd_model.device) return sampling.append_zero(self.t_to_sigma(t)) def sigma_to_t(self, sigma, quantize=None): log_sigma = sigma.log() dists = log_sigma - self.log_sigmas[:, None] return dists.abs().argmin(dim=0).view(sigma.shape) * self.skip_steps + (self.skip_steps - 1) def t_to_sigma(self, timestep): t = torch.clamp(((timestep - (self.skip_steps - 1)) / self.skip_steps).float(), min=0, max=(len(self.sigmas) - 1)) return super().t_to_sigma(t) def get_eps(self, *args, **kwargs): return self.inner_model.apply_model(*args, **kwargs) def get_scaled_out(self, sigma, output, input): sigma_data = 0.5 scaled_timestep = utils.append_dims(self.sigma_to_t(sigma), output.ndim) * 10.0 c_skip = sigma_data**2 / (scaled_timestep**2 + sigma_data**2) c_out = scaled_timestep / (scaled_timestep**2 + sigma_data**2) ** 0.5 return c_out * output + c_skip * input def forward(self, input, sigma, **kwargs): c_out, c_in = [utils.append_dims(x, input.ndim) for x in self.get_scalings(sigma)] eps = self.get_eps(input * c_in, self.sigma_to_t(sigma), **kwargs) return self.get_scaled_out(sigma, input + eps * c_out, input) class CFGDenoiserLCM(sd_samplers_cfg_denoiser.CFGDenoiser): @property def inner_model(self): if self.model_wrap is None: denoiser = LCMCompVisDenoiser self.model_wrap = denoiser(shared.sd_model) return self.model_wrap

 

webuiを再起動してリロードすると追加されます。

サンプリングメソッドにRestartLabが三つ追加されます。

 

設定用UIもエクステンションの所に追加されます。

 

設定項目

基本設定

first weight

最初のステップ数をコントロールできます。restart stepを掛けた値が大体のステップ数になります。

掛けた値がステップ数を超える時は、1で処理されます。

 

restart Step

リスタートする間隔を設定します。ステップ数の半分以下にしてください。

リスタートしたくない時は、大きい値を設定してください。

 

CFG Denoiser

デフォルトのDiffusionとLCMが選択できます。

 

Sampling

Heun 、Euler系、DPM++系が選択できます。

 

Noise

SDE系用のSDE、Euler a用のAncestral 、Brownianが選べます。

Brownian remakeは、リスタートする時にノイズを作成し直します。

ノイズが影響するのは、Euler aとSDE系です。 

設定の初期値は、SetingsのRestartLabで設定できます。

 

Sampler Parameters

CFG、s_churn、s_noiseの量をステップごとにコントロールできます。

start s_churn、s_noise
最初の値です。

Change CFG、End s_churn、s_noise
途中で変える値です。
CFGは、大きく変化すると破綻しやすいのでコントロールは難しそうです。

step
値を変えたいステップ数です。１以下の場合はステップ数の割合で指定できます。
0の場合は何もしません。

mode
変化の動作を設定できます。
OFF：処理しません。
Constant：startの値で固定します。
After：v1.0の動作と同じです。設定したステップ数の次のステップから変更します。
Linear：直線で変化します。
Slow：ゆっくり変化してだんだん早くなります。
Fast：早く変化してだんだんゆっくりになります。
Turn：Fastの動作と同じですが、最初の値に戻る動作をします。

数字の違いは、2は、なだらかな動作、６は、極端な動作をします。４はその中間です。

AutoSkip

途中で生成を止めて、止めた所から再開する事が出来ます。
プロンプトの調整や、構図を探す時に使っています。
※Batch sizeは、１以外だと動作しないかもしれません。
※Hires.fixも使ってないので動作未確認です。

SkipMode
％はステップ数の割合です。60%をよく使います。
Customは、CustomStepで止めるステップを直接設定したい時に使います。
Nextは、Restart毎に止めたい時に、
1st、2ndなどは、止めたいRestartの回数です。1stは一回目、2ndは2回目で止まります。

その他の設定
Save SkipMeta to pinginfo：生成途中のデータをpnginfoに保存します。大きい画像だとpnginfoが大きくなるのでオフの方が良いかもしれません。オフの場合は、再開できません。
Use SkipMeta：SkipMetaを使用して途中から生成します。アップスケールする時は、オフにしてください。
Make the first step SkipMeta：Use SkipMetaがオン時、最初のステップにSkipMetaを使用します。VAEを通してないデータでi2iする感じです。たぶん 　最後にVAE使ってると思ってたけど、ステップごとに使ってるのかも。なので言い換えると画像にする前のデータでi2iする感じです。

SkipMeta：途中のデータです。

Batch

再開したり、アップスケールする時に使います。
Batch Processに処理したいファイルをドロップします。
Parameters to remove from png infoからpnginfoから、pnginfoから読み込みたくない項目を選択します。
※アップスケールする時に、StepsとDenoising strenghにチェックしたりします。

Startボタンで開始します。
※オートクリックの様な処理方法なのでブラウザは開いたままにしてください。
※あと、たまに、pnginfoがsend出来なかったりで処理が止まる時があります。

Tools

Batchで使ってるボタン等です。ここのボタンを自動でクリックしているので通常は使いません。
たまに、sourceにpnginfoをsendしたいファイルを入れて、send to img2imgボタンを使います。

参考画像

<lora:pytorch_lora_weights:0.6>

Steps: 12, Sampler: RestartLab karras, CFG scale: 2

RestartLab firstweight: 2, RestartLab cfgDenoiser: CFGDenoiserKDiffusion

steps 8は、12ステップなのでリスタートしてないです。

steps 5は、一回リスタートしています。

 

<lora:pytorch_lora_weights:0.6>

Steps: 9, Sampler: RestartLab karras, CFG scale: 2,

RestartLab steps: 4, RestartLab firstweight: 1, RestartLab cfgDenoiser: CFGDenoiserKDiffusion, RestartLab sampling: Euler a

ノイズの違いです。一回リセットしてます。