73 lines
3.1 KiB
Python
73 lines
3.1 KiB
Python
import torch
|
|
|
|
class uncondZeroNode:
|
|
@classmethod
|
|
def INPUT_TYPES(s):
|
|
return {"required": {
|
|
"model": ("MODEL",),
|
|
"scale": ("FLOAT", {"default": 1, "min": 0.0, "max": 10.0, "step": 0.01, "round": 0.01}),
|
|
"method":(["uncond_zero_v1","uncond_zero_v2","uncond_zero_v3"], {"default": "uncond_zero_v3"},),
|
|
}}
|
|
RETURN_TYPES = ("MODEL",)
|
|
FUNCTION = "patch"
|
|
|
|
CATEGORY = "model_patches"
|
|
|
|
def patch(self, model, scale, method):
|
|
def uncond_zero(args):
|
|
cond = args["cond_denoised"]
|
|
x_orig = args["input"]
|
|
x_orig -= x_orig.mean() # the main trick to not get a mess is simply to subtract the mean values. I guess SD likes it gaussian AF
|
|
cond -= cond.mean()
|
|
return x_orig - cond / cond.std() ** .5 * scale # the square root of the std is simply an ever changing scale that fits the bill. The only true condition is to have it not above one near the end.
|
|
|
|
def uncond_zero_v2(args):
|
|
cond = args["cond_denoised"]
|
|
x_orig = args["input"]
|
|
cond -= cond.mean()
|
|
result = torch.zeros_like(x_orig)
|
|
|
|
for b in range(len(x_orig)):
|
|
for c in range(len(x_orig[b])):
|
|
x_orig[b][c] -= x_orig[b][c].mean()
|
|
cond_c_mean = cond[b][c].mean()
|
|
cond[b][c] -= cond_c_mean
|
|
result[b][c] = x_orig[b][c] - cond[b][c] / cond[b][c].std() ** .5 * scale + cond_c_mean
|
|
|
|
return result
|
|
|
|
|
|
# new_scale = 1 / (model.model.latent_format.scale_factor * 8) # Anything below this value gave visible artifacts.
|
|
# alright it was around 0.95 with SDXL and 1 is just better. SD 1.x latent scale gives a lower value which ended in bad results.
|
|
|
|
#Taken and modified from comfy_extras/nodes_model_advanced
|
|
def rescale_cfg(args):
|
|
x_orig = args["input"]
|
|
x_orig -= x_orig.mean()
|
|
|
|
cond = args["cond_denoised"]
|
|
cond -= cond.mean()
|
|
|
|
cond = x_orig - cond / cond.std() ** .5 * min(scale, 1)
|
|
|
|
sigma = args["sigma"]
|
|
sigma = sigma.view(sigma.shape[:1] + (1,) * (cond.ndim - 1))
|
|
|
|
#rescale cfg has to be done on v-pred model output
|
|
x = x_orig / (sigma * sigma + 1.0)
|
|
uncond = x / sigma
|
|
cond = ((x - (x_orig - cond)) * (sigma ** 2 + 1.0) ** 0.5) / (sigma)
|
|
|
|
#rescalecfg
|
|
x_cfg = uncond + max(scale, 1) * (cond - uncond)
|
|
ro_pos = torch.std(cond, dim=(1,2,3), keepdim=True)
|
|
ro_cfg = torch.std(x_cfg, dim=(1,2,3), keepdim=True)
|
|
|
|
x_rescaled = x_cfg * (ro_pos / ro_cfg)
|
|
|
|
return x_orig - (x - x_rescaled * sigma / (sigma * sigma + 1.0) ** 0.5)
|
|
|
|
m = model.clone()
|
|
m.set_model_sampler_cfg_function({"uncond_zero_v1":uncond_zero,"uncond_zero_v2":uncond_zero_v2,"uncond_zero_v3":rescale_cfg}[method])
|
|
return (m, )
|