laksjdjf’s gists

laksjdjf / visual_style_prompting.py

Created February 28, 2024 13:34

test of https://curryjung.github.io/VisualStylePrompt/

	import torch

	class VisualStylePrompting:
	@classmethod
	def INPUT_TYPES(s):
	return {"required": {
	"model": ("MODEL",),
	"reference": ("LATENT",),
	"depth": ("INT", {"default": 0, "min": -1, "max": 12}),
	"batch_size": ("INT", {"default": 1, "min": 1, "max": 64}),

laksjdjf / scale_crafter.py

Created February 24, 2024 17:04

https://github.com/YingqingHe/ScaleCrafter

	# ref: ScaleCrafter https://github.com/YingqingHe/ScaleCrafter

	import math
	import comfy.ops
	import torch.nn.functional as F
	ops = comfy.ops.disable_weight_init

	class ScaleCrafter:
	@classmethod
	def INPUT_TYPES(s):

laksjdjf / dilate_conv.py

Last active March 6, 2024 11:35

Reference from ScaleCrafter[https://arxiv.org/abs/2310.07702]

	# https://arxiv.org/abs/2310.07702

	import comfy.ops
	ops = comfy.ops.disable_weight_init

	class DilateConv:
	@classmethod
	def INPUT_TYPES(s):
	return {
	"required": {

laksjdjf / stable_cascade_decoder.txt

Created February 14, 2024 13:37

	===========================================================================================================================================================
	Layer (type (var_name)) Input Shape Output Shape Param # Kernel Shape
	===========================================================================================================================================================
	StableCascadeUnet (StableCascadeUnet) -- [1, 4, 256, 256] -- 3
	├─Linear (clip_txt_pooled_mapper) [1, 1, 1280] [1, 1, 5120] 6,558,720 --
	├─LayerNorm (clip_norm) [1, 4, 1280] [1, 4, 1280] -- --
	├─Sequential (embedding) [1, 4, 256, 256] [1, 320, 128, 128] -- --
	│ └─PixelUnshuf

laksjdjf / stable_cascade_prior.txt

Created February 14, 2024 13:36

	===========================================================================================================================================================
	Layer (type (var_name)) Input Shape Output Shape Param # Kernel Shape
	===========================================================================================================================================================
	StableCascadeUnet (StableCascadeUnet) [2, 16, 24, 24] [2, 16, 24, 24] 8,923,136 3
	├─Linear (clip_txt_pooled_mapper) [2, 77, 1280] [2, 77, 8192] 10,493,952 --
	├─LayerNorm (clip_norm) [2, 308, 2048] [2, 308, 2048] -- --
	├─Sequential (embedding) [2, 16, 24, 24] [2, 2048, 24, 24] -- --
	│ └─PixelUnshuf

laksjdjf / text_scheduler.py

Created February 12, 2024 08:15

	'''
	load from sampling/custom_sampling/scheulers
	input text like "999,893,...,156"
	connect to SamplerCustom
	'''

	import torch

	class TextScheduler:
	@classmethod

laksjdjf / LCMSamplerRCFG.py

Last active December 21, 2023 12:02

Implementation of RCFG in https://arxiv.org/abs/2312.12491

	'''
	Implementation of RCFG in https://arxiv.org/abs/2312.12491
	Node is in sampling/custom_sampling/samplers
	original_latent is OPTIONAL
	If original_latent is set, it is Self-Negative else Onetime-Negative
	cfg is recommendet near 1.0 (KSAMPLER"s cfg is ignored)
	delta is よくわかんない
	'''

	from comfy.samplers import KSAMPLER

laksjdjf / DeepCache.py

Last active February 17, 2025 06:12

https://arxiv.org/abs/2312.00858

	'''
	https://arxiv.org/abs/2312.00858
	1. put this file in ComfyUI/custom_nodes
	2. load node from <loaders>

	start_step, end_step: apply this method when the timestep is between start_step and end_step
	cache_interval: interval of caching (1 means no caching)
	cache_depth: depth of caching
	'''

laksjdjf / gradual_latent.py

Last active December 1, 2023 13:36 — forked from kohya-ss/gradual_latent.py

	import comfy
	from comfy.samplers import KSAMPLER
	import torch
	from torchvision.transforms.functional import gaussian_blur
	from comfy.k_diffusion.sampling import default_noise_sampler, get_ancestral_step, to_d, BrownianTreeNoiseSampler
	from tqdm.auto import trange

	def interpolate(x, size, unsharp_strength=0.0, unsharp_kernel_size=3, unsharp_sigma=0.5, unsharp=False, mode="bicubic", align_corners=False):
	x = torch.nn.functional.interpolate(x, size=size, mode=mode, align_corners=align_corners)
	if unsharp_strength > 0 and unsharp:

laksjdjf / gradual_latent.py

Last active November 30, 2023 23:27

	import comfy
	from comfy.samplers import KSAMPLER
	import torch
	from comfy.k_diffusion.sampling import default_noise_sampler, get_ancestral_step, to_d, BrownianTreeNoiseSampler
	from tqdm.auto import trange

	@torch.no_grad()
	def sample_euler_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, upscale_ratio=2.0, start_step=5, end_step=15, upscale_n_step=3):
	"""Ancestral sampling with Euler method steps."""
	extra_args = {} if extra_args is None else extra_args