310 lines
14 KiB
Python
310 lines
14 KiB
Python
import hashlib
|
|
import math
|
|
import random
|
|
import re
|
|
import torch
|
|
import torch.nn.functional as F
|
|
|
|
import comfy.sample
|
|
import comfy.samplers
|
|
import comfy.model_management
|
|
import nodes
|
|
|
|
def slerp(strength, tensor_from, tensor_to, epsilon=1e-6):
|
|
"""
|
|
Perform Spherical Linear Interpolation (Slerp) between two tensors.
|
|
|
|
Parameters:
|
|
- strength (float): The interpolation factor between tensor_from and tensor_to.
|
|
- tensor_from (Tensor): The starting tensor.
|
|
- tensor_to (Tensor): The ending tensor.
|
|
- epsilon (float): division by zero offset
|
|
|
|
Returns:
|
|
- Tensor: Interpolated tensor.
|
|
"""
|
|
low_norm = F.normalize(tensor_from, p=2, dim=-1, eps=epsilon)
|
|
high_norm = F.normalize(tensor_to, p=2, dim=-1, eps=epsilon)
|
|
|
|
dot_product = torch.clamp((low_norm * high_norm).sum(dim=-1), -1.0, 1.0)
|
|
omega = torch.acos(dot_product)
|
|
so = torch.sin(omega)
|
|
zero_so_mask = torch.isclose(so, torch.tensor([0.0], device=so.device), atol=epsilon)
|
|
so = torch.where(zero_so_mask, torch.tensor([1.0], device=so.device), so)
|
|
sin_omega_minus_strength = torch.sin((1.0 - strength) * omega) / so
|
|
sin_strength_omega = torch.sin(strength * omega) / so
|
|
|
|
res = sin_omega_minus_strength.unsqueeze(-1) * tensor_from + sin_strength_omega.unsqueeze(-1) * tensor_to
|
|
res = torch.where(zero_so_mask.unsqueeze(-1),
|
|
tensor_from if strength < 0.5 else tensor_to,
|
|
res)
|
|
|
|
return res
|
|
|
|
# from https://discuss.pytorch.org/t/help-regarding-slerp-function-for-generative-model-sampling/32475
|
|
def slerp_latents(val, low, high):
|
|
dims = low.shape
|
|
|
|
#flatten to batches
|
|
low = low.reshape(dims[0], -1)
|
|
high = high.reshape(dims[0], -1)
|
|
|
|
low_norm = low/torch.norm(low, dim=1, keepdim=True)
|
|
high_norm = high/torch.norm(high, dim=1, keepdim=True)
|
|
|
|
# in case we divide by zero
|
|
low_norm[low_norm != low_norm] = 0.0
|
|
high_norm[high_norm != high_norm] = 0.0
|
|
|
|
omega = torch.acos((low_norm*high_norm).sum(1))
|
|
so = torch.sin(omega)
|
|
res = (torch.sin((1.0-val)*omega)/so).unsqueeze(1)*low + (torch.sin(val*omega)/so).unsqueeze(1) * high
|
|
return res.reshape(dims)
|
|
|
|
def blend_latents(alpha, latent_1, latent_2):
|
|
if not isinstance(alpha, torch.Tensor):
|
|
alpha = torch.tensor([alpha], dtype=latent_1.dtype, device=latent_1.device)
|
|
|
|
blended_latent = (1 - alpha) * latent_1 + alpha * latent_2
|
|
|
|
return blended_latent
|
|
|
|
def cosine_interp_latents(val, low, high):
|
|
if not isinstance(val, torch.Tensor):
|
|
val = torch.tensor([val], dtype=low.dtype, device=low.device)
|
|
t = (1 - torch.cos(val * math.pi)) / 2
|
|
return (1 - t) * low + t * high
|
|
|
|
def unsample(model, seed, cfg, sampler_name, steps, end_at_step, scheduler, normalize, positive, negative, latent_image):
|
|
device = comfy.model_management.get_torch_device()
|
|
end_at_step = steps - min(end_at_step, steps - 1)
|
|
|
|
latent = latent_image
|
|
latent_image = latent["samples"].to(device)
|
|
|
|
noise_shape = latent_image.size()
|
|
noise = torch.zeros(noise_shape, dtype=latent_image.dtype, layout=latent_image.layout, device=device)
|
|
noise_mask = comfy.sample.prepare_mask(latent.get("noise_mask"), noise, device) if "noise_mask" in latent else None
|
|
|
|
positive_copy = comfy.sample.broadcast_cond(positive, noise.shape[0], device)
|
|
negative_copy = comfy.sample.broadcast_cond(negative, noise.shape[0], device)
|
|
|
|
models, inference_memory = comfy.sample.get_additional_models(positive, negative, model.model_dtype())
|
|
comfy.model_management.load_models_gpu([model] + models, comfy.model_management.batch_area_memory(noise.numel() // noise.shape[0]) + inference_memory)
|
|
|
|
real_model = model.model
|
|
sampler = comfy.samplers.KSampler(real_model, steps=steps, device=device, sampler=sampler_name, scheduler=scheduler, denoise=1.0, model_options=model.model_options)
|
|
sigmas = sampler.sigmas.flip(0) + 0.0001
|
|
|
|
pbar = comfy.utils.ProgressBar(steps)
|
|
def callback(step, x0, x, total_steps): pbar.update_absolute(step + 1, total_steps)
|
|
|
|
samples = sampler.sample(noise, positive_copy, negative_copy, cfg=cfg, latent_image=latent_image, force_full_denoise=False, denoise_mask=noise_mask, sigmas=sigmas, start_step=0, last_step=end_at_step, callback=callback, seed=seed)
|
|
|
|
if normalize == "enable":
|
|
samples = (samples - samples.mean()) / samples.std()
|
|
|
|
comfy.sample.cleanup_additional_models(models)
|
|
|
|
out = latent.copy()
|
|
out["samples"] = samples.cpu()
|
|
return (out,)
|
|
|
|
CLIPTextEncode = nodes.CLIPTextEncode()
|
|
USE_BLK, BLK_ADV = (False, None)
|
|
if "BNK_CLIPTextEncodeAdvanced" in nodes.NODE_CLASS_MAPPINGS:
|
|
BLK_ADV = nodes.NODE_CLASS_MAPPINGS['BNK_CLIPTextEncodeAdvanced']
|
|
USE_BLK = True
|
|
|
|
if USE_BLK:
|
|
print(f"Found `\33[1mComfyUI_ADV_CLIP_emb\33[0m`. Using \33[93mBLK Advanced CLIPTextEncode\33[0m for Conditioning Sequencing")
|
|
blk_adv = BLK_ADV()
|
|
|
|
class CLIPTextEncodeSequence:
|
|
@classmethod
|
|
def INPUT_TYPES(s):
|
|
return {
|
|
"required": {
|
|
"clip": ("CLIP", ),
|
|
"token_normalization": (["none", "mean", "length", "length+mean"],),
|
|
"weight_interpretation": (["comfy", "A1111", "compel", "comfy++"],),
|
|
"text": ("STRING", {"multiline": True, "default": '''0:A portrait of a rosebud
|
|
5:A portrait of a blooming rosebud
|
|
10:A portrait of a blooming rose
|
|
15:A portrait of a rose'''}),
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("CONDITIONING_SEQ",)
|
|
RETURN_NAMES = ("conditioning_sequence",)
|
|
IS_LIST_OUTPUT = (True,)
|
|
|
|
FUNCTION = "encode"
|
|
CATEGORY = "conditioning"
|
|
|
|
def encode(self, clip, text, token_normalization, weight_interpretation):
|
|
text = text.strip()
|
|
conditionings = []
|
|
for l in text.splitlines():
|
|
match = re.match(r'(\d+):', l)
|
|
|
|
if match:
|
|
idx = int(match.group(1))
|
|
_, line = l.split(":", 1)
|
|
line = line.strip()
|
|
|
|
if USE_BLK:
|
|
encoded = blk_adv.encode(clip=clip, text=line, token_normalization=token_normalization, weight_interpretation=weight_interpretation)
|
|
else:
|
|
encoded = CLIPTextEncode.encode(clip=clip, text=line)
|
|
|
|
conditioning = (idx, [encoded[0][0][0], encoded[0][0][1]])
|
|
conditionings.append(conditioning)
|
|
|
|
return (conditionings, )
|
|
|
|
class KSamplerSeq:
|
|
@classmethod
|
|
def INPUT_TYPES(s):
|
|
return {"required":
|
|
{"model": ("MODEL",),
|
|
"seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
|
|
"seed_mode_seq": (["increment", "decrement", "random", "fixed"],),
|
|
"steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
|
|
"cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.5, "round": 0.01}),
|
|
"sampler_name": (comfy.samplers.KSampler.SAMPLERS, ),
|
|
"scheduler": (comfy.samplers.KSampler.SCHEDULERS, ),
|
|
"sequence_loop_count": ("INT", {"default": 20, "min": 1, "max": 100, "step": 1}),
|
|
"positive_seq": ("CONDITIONING_SEQ", ),
|
|
"negative_seq": ("CONDITIONING_SEQ", ),
|
|
"use_conditioning_slerp": ("BOOLEAN", {"default": False}),
|
|
"cond_slerp_strength": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.001}),
|
|
"latent_image": ("LATENT", ),
|
|
"use_latent_interpolation": ("BOOLEAN", {"default": False}),
|
|
"latent_interpolation_mode": (["Blend", "Slerp", "Cosine Interp"],),
|
|
"latent_interp_strength": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.001}),
|
|
"denoise_start": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
|
|
"denoise_seq": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}),
|
|
"unsample_latents": ("BOOLEAN", {"default": False})
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("LATENT",)
|
|
FUNCTION = "sample"
|
|
|
|
CATEGORY = "sampling"
|
|
|
|
def update_seed(self, seed, seed_mode):
|
|
if seed_mode == "increment":
|
|
return seed + 1
|
|
elif seed_mode == "decrement":
|
|
return seed - 1
|
|
elif seed_mode == "random":
|
|
return random.randint(0, 0xffffffffffffffff)
|
|
elif seed_mode == "fixed":
|
|
return seed
|
|
|
|
def hash_tensor(self, tensor):
|
|
tensor = tensor.cpu().contiguous()
|
|
return hashlib.sha256(tensor.numpy().tobytes()).hexdigest()
|
|
|
|
def update_conditioning(self, conditioning_seq, loop_count, last_conditioning):
|
|
matching_conditioning = None
|
|
for idx, conditioning, *_ in conditioning_seq:
|
|
if int(idx) == loop_count:
|
|
matching_conditioning = conditioning
|
|
break
|
|
return matching_conditioning if matching_conditioning else (last_conditioning if last_conditioning else None)
|
|
|
|
def sample(self, model, seed, seed_mode_seq, steps, cfg, sampler_name, scheduler, sequence_loop_count, positive_seq, negative_seq, cond_slerp_strength, latent_image, use_latent_interpolation, latent_interpolation_mode, latent_interp_strength, denoise_start=1.0, denoise_seq=0.5, use_conditioning_slerp=False, unsample_latents=False):
|
|
positive_seq = positive_seq
|
|
negative_seq = negative_seq
|
|
|
|
results = []
|
|
positive_conditioning = None
|
|
negative_conditioning = None
|
|
for loop_count in range(sequence_loop_count):
|
|
|
|
seq_seed = seed if loop_count <= 0 else self.update_seed(seq_seed, seed_mode_seq)
|
|
|
|
print(f"Loop count: {loop_count}, Seed: {seq_seed}")
|
|
|
|
last_positive_conditioning = positive_conditioning[0] if positive_conditioning else None
|
|
last_negative_conditioning = negative_conditioning[0] if negative_conditioning else None
|
|
|
|
positive_conditioning = self.update_conditioning(positive_seq, loop_count, last_positive_conditioning)
|
|
negative_conditioning = self.update_conditioning(negative_seq, loop_count, last_negative_conditioning)
|
|
|
|
if use_conditioning_slerp and (last_positive_conditioning and last_negative_conditioning):
|
|
a = last_positive_conditioning[0].clone()
|
|
b = positive_conditioning[0].clone()
|
|
na = last_negative_conditioning[0].clone()
|
|
nb = negative_conditioning[0].clone()
|
|
|
|
pa = last_positive_conditioning[1]["pooled_output"].clone()
|
|
pb = positive_conditioning[1]["pooled_output"].clone()
|
|
npa = last_negative_conditioning[1]["pooled_output"].clone()
|
|
npb = negative_conditioning[1]["pooled_output"].clone()
|
|
|
|
pos_cond = slerp(cond_slerp_strength, a, b)
|
|
pos_pooled = slerp(cond_slerp_strength, pa, pb)
|
|
neg_cond = slerp(cond_slerp_strength, na, nb)
|
|
neg_pooled = slerp(cond_slerp_strength, npa, npb)
|
|
|
|
positive_conditioning = [pos_cond, {"pooled_output": pos_pooled}]
|
|
negative_conditioning = [neg_cond, {"pooled_output": neg_pooled}]
|
|
|
|
positive_conditioning = [positive_conditioning]
|
|
negative_conditioning = [negative_conditioning]
|
|
|
|
if positive_conditioning is not None or negative_conditioning is not None:
|
|
|
|
end_at_step = steps
|
|
if results is not None and len(results) > 0:
|
|
latent_input = {'samples': results[-1]}
|
|
denoise = denoise_seq
|
|
start_at_step = round((1 - denoise) * steps)
|
|
end_at_step = steps
|
|
else:
|
|
latent_input = latent_image
|
|
denoise = denoise_start
|
|
|
|
if unsample_latents and loop_count > 0:
|
|
force_full_denoise = False if loop_count > 0 or loop_count <= steps - 1 else True
|
|
disable_noise = False
|
|
unsampled_latent = unsample(model=model, seed=seq_seed, cfg=cfg, sampler_name=sampler_name, steps=steps, end_at_step=end_at_step, scheduler=scheduler, normalize=False, positive=positive_conditioning, negative=negative_conditioning, latent_image=latent_input)[0]
|
|
sample = nodes.common_ksampler(model, seq_seed, steps, cfg, sampler_name, scheduler, positive_conditioning, negative_conditioning, unsampled_latent, denoise=denoise, disable_noise=disable_noise, start_step=start_at_step, last_step=end_at_step, force_full_denoise=force_full_denoise)[0]['samples']
|
|
else:
|
|
sample = nodes.common_ksampler(model, seq_seed, steps, cfg, sampler_name, scheduler, positive_conditioning, negative_conditioning, latent_input, denoise=denoise)[0]['samples']
|
|
|
|
if use_latent_interpolation and results and loop_count > 0:
|
|
if latent_interpolation_mode == "Blend":
|
|
sample = blend_latents(latent_interp_strength, results[-1], sample)
|
|
elif latent_interpolation_mode == "Slerp":
|
|
sample = slerp_latents(latent_interp_strength, results[-1], sample)
|
|
elif latent_interpolation_mode == "Cosine Interp":
|
|
sample = cosine_interp_latents(latent_interp_strength, results[-1], sample)
|
|
else:
|
|
sample = sample
|
|
|
|
results.append(sample)
|
|
|
|
results = torch.cat(results, dim=0)
|
|
results = {'samples': results}
|
|
|
|
return (results,)
|
|
|
|
|
|
|
|
|
|
NODE_CLASS_MAPPINGS = {
|
|
"CLIPTextEncodeList": CLIPTextEncodeSequence,
|
|
"KSamplerSeq": KSamplerSeq,
|
|
}
|
|
|
|
NODE_DISPLAY_NAME_MAPPINGS = {
|
|
"CLIPTextEncodeList": "CLIP Text Encode Sequence (Advanced)",
|
|
"KSamplerSeq": "KSampler Sequence",
|
|
|
|
}
|