blepping
d8574bae2d
* Weird experiments with sampler blending * Mega sync * Use ComfyUI union types for wildcard inputs when available * Add weoon wavelet sampler * Fix noise sampler caching not considering immiscible settings Allow specifying alt custom noise/immiscible settings for samplers that internally add noise AFS support at the merge sampler level Allow the Weoon internal step to use ETA Add t_copysign expression function * Start updating documentation * Add ImmiscibleReference node, other changes * Respect factor in ImmiscibleReference noise * Fixing normalizing in ImmiscibleReference noise * Handle case where there are no RGB factors (i.e. audio models) * Fix rectified flow type detection for non-Flux models * Add more tensor ops, add OCS ApplyExpressionImage/Latent nodes * Rename ApplyExpression nodes to ApplyFilter, expression QoL improvements * Fix apply filter nodes definitions * Make t_noise handler work for image-latents. * Add t_new_like expression function * Make it possible to comment out lines with hash in expressions * Add gradient estimation, pingpong and res_multistep samplers. Add pingpong group merge method. Remove model call caching stuff. Allow defining pre_cfg and post_cfg filters. Handle filters changing the latent shape better. Allow disabling detecting out of order sigmas as restart sampling. * Sync current updates (which I am too lazy to describe individually) * Add ExpressionFilteredNoise node Improvements to immiscible noise handling/expanded features
30 lines
930 B
Python
30 lines
930 B
Python
from .base import SingleStepSampler, registry
|
|
|
|
|
|
# Referenced from https://github.com/ace-step/ACE-Step/
|
|
class PingPongStep(SingleStepSampler):
|
|
name = "pingpong"
|
|
default_eta = 0.0
|
|
|
|
def __init__(self, *args, **kwargs):
|
|
super().__init__(*args, **kwargs)
|
|
pingpong_options = self.options.pop("pingpong", {})
|
|
self.pingpong_start_step = pingpong_options.get("start_step", 0)
|
|
self.pingpong_end_step = pingpong_options.get("end_step", 0)
|
|
|
|
def step(self, x):
|
|
ss = self.ss
|
|
use_pingpong = self.pingpong_start_step <= ss.step <= self.pingpong_end_step
|
|
if not use_pingpong:
|
|
return (yield from self.euler_step(x, eta=0.0))
|
|
sn = ss.sigma_next
|
|
denoised = (
|
|
ss.denoised * (1.0 - sn) if ss.model.is_rectified_flow else ss.denoised
|
|
)
|
|
yield from self.result(denoised, sn, sigma_down=sn)
|
|
|
|
|
|
registry.add(
|
|
PingPongStep,
|
|
)
|