401 lines
19 KiB
Python
401 lines
19 KiB
Python
"""
|
|
ComfyUI节点 - 动态参数输入
|
|
支持独立选择 sampler_method 和 model_method，可单独或同时使用
|
|
"""
|
|
|
|
import comfy
|
|
|
|
support_sampler = ["euler", "heun", "heunpp2","dpm_2","lms", "dpm_fast", "dpm_adaptive",
|
|
"dpmpp_2m", "ipndm", "ipndm_v", "deis", "res_multistep",
|
|
"gradient_estimation","ddim", "euler_ancestral","dpm_2_ancestral"]
|
|
|
|
# SADA supports a subset of samplers
|
|
SADA_SUPPORT_SAMPLER = ["euler", "heun", "heunpp2", "dpm_2", "lms", "dpmpp_2m", "ddim",
|
|
"euler_ancestral", "dpm_2_ancestral"]
|
|
|
|
TEACACHE_MODEL_TYPES = [
|
|
"flux", "flux-kontext", "ltxv", "lumina_2", "hunyuan_video",
|
|
"hidream_i1_full", "hidream_i1_dev", "hidream_i1_fast",
|
|
"wan2.1_t2v_1.3B", "wan2.1_t2v_14B",
|
|
"wan2.1_i2v_480p_14B", "wan2.1_i2v_720p_14B",
|
|
"wan2.1_t2v_1.3B_ret_mode", "wan2.1_t2v_14B_ret_mode",
|
|
"wan2.1_i2v_480p_14B_ret_mode", "wan2.1_i2v_720p_14B_ret_mode",
|
|
]
|
|
|
|
MAGCACHE_MODEL_TYPES = [
|
|
"flux", "flux_kontext", "chroma", "qwen_image",
|
|
"hunyuan_video", "hunyuan_video1.5",
|
|
"wan2.1_t2v_1.3B", "wan2.1_t2v_14B",
|
|
"wan2.1_i2v_480p_14B", "wan2.1_i2v_720p_14B",
|
|
"wan2.1_vace_1.3B", "wan2.1_vace_14B",
|
|
]
|
|
|
|
TAYLORSEER_MODEL_TYPES = ["flux", "hunyuan_video", "wan2.1", "hidream"]
|
|
|
|
HICACHE_MODEL_TYPES = ["flux", "hunyuan_video", "wan2.1", "hidream"]
|
|
HICACHE_PREDICTION_MODES = ["hicache", "taylor_scaled"]
|
|
|
|
SEACACHE_MODEL_TYPES = ["flux", "hunyuan_video", "wan2.1"]
|
|
|
|
TCPADE_MODEL_TYPES = ["flux", "wan2.1"]
|
|
|
|
# ========== sampler_method 可选项 ==========
|
|
SAMPLER_METHOD_LIST = ["None", "AdaptiveDiff", "EasyCache", "SADA", "ZEUS"]
|
|
|
|
ZEUS_INTERP_MODES = ["psi", "x_0"]
|
|
ZEUS_CACHING_MODES = ["reuse_interp", "interp_all", "reuse_all"]
|
|
|
|
# ========== model_method (model-feature) 可选项 ==========
|
|
MODEL_METHOD_LIST = ["None", "TeaCache", "MagCache", "TaylorSeer", "HiCache", "SeaCache", "TC-Pade"]
|
|
|
|
|
|
def _apply_teacache(model, model_type, rel_l1_thresh, start_percent, end_percent, cache_device):
|
|
from .model.teacache.nodes import TeaCache
|
|
return TeaCache().apply_teacache(
|
|
model, model_type, rel_l1_thresh, start_percent, end_percent, cache_device
|
|
)[0]
|
|
|
|
|
|
def _apply_taylorseer(model, model_type, max_order, fresh_threshold, first_enhance):
|
|
from .model.taylorseer.nodes import apply_taylorseer
|
|
return apply_taylorseer(model, model_type, max_order, fresh_threshold, first_enhance)
|
|
|
|
|
|
def _apply_hicache(model, model_type, max_order, fresh_threshold, first_enhance,
|
|
prediction_mode, hicache_scale_factor):
|
|
from .model.taylorseer.nodes import apply_taylorseer
|
|
return apply_taylorseer(model, model_type, max_order, fresh_threshold, first_enhance,
|
|
prediction_mode=prediction_mode,
|
|
hicache_scale_factor=hicache_scale_factor)
|
|
|
|
|
|
def _apply_seacache(model, model_type, seacache_thresh, power_exp, ret_steps):
|
|
from .model.seacache.nodes import apply_seacache
|
|
return apply_seacache(model, model_type, seacache_thresh, power_exp, ret_steps)
|
|
|
|
|
|
def _apply_tcpade(model, model_type, start_step, end_step, interval, n_threshold,
|
|
predictor_order, history_size, cache_device):
|
|
from .model.tcpade.nodes import apply_tcpade
|
|
return apply_tcpade(
|
|
model, model_type, start_step, end_step, interval, n_threshold,
|
|
predictor_order, history_size, cache_device,
|
|
)
|
|
|
|
|
|
def _apply_magcache(model, model_type, magcache_thresh, retention_ratio, magcache_K, start_step, end_step):
|
|
from .model.magcache.nodes import MagCache
|
|
return MagCache().apply_magcache(
|
|
model, model_type, magcache_thresh, retention_ratio, magcache_K, start_step, end_step
|
|
)[0]
|
|
|
|
|
|
|
|
class DynamicModelParamsNode:
|
|
"""
|
|
支持独立选择 sampler_method 和 model_method 的统一加速节点。
|
|
- sampler_method: 选择 sampler 加速方法（SADA, AdaptiveDiff, EasyCache, ZEUS），None 表示不使用
|
|
- model_method: 选择 model-feature 加速方法（TeaCache, MagCache 等），None 表示不使用
|
|
- 两者可独立使用，也可同时使用
|
|
"""
|
|
|
|
def __init__(self):
|
|
pass
|
|
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"sampler_method": (SAMPLER_METHOD_LIST, {"default": "None"}),
|
|
"model_method": (MODEL_METHOD_LIST, {"default": "None"}),
|
|
},
|
|
"optional": {
|
|
# 具体的采样器选择 (euler, heun, etc.)
|
|
"sampler_name": (support_sampler,),
|
|
# MODEL 输入
|
|
"model": ("MODEL",),
|
|
# === AdaptiveDiff sampler 参数 ===
|
|
"max_skip_steps": ("INT", {"default": 4, "min": 0, "max": 10, "step": 1}),
|
|
"threshold": ("FLOAT", {"default": 0.01, "min": 0.0, "max": 1.0, "step": 0.001}),
|
|
# === EasyCache sampler 参数 ===
|
|
"ec_threshold": ("FLOAT", {"default": 0.025, "min": 0.0, "max": 1.0, "step": 0.001}),
|
|
"ret_steps": ("INT", {"default": 5, "min": 0, "max": 100, "step": 1}),
|
|
# === SADA sampler 参数 ===
|
|
"lagrange_term": ("INT", {"default": 3, "min": 0, "max": 5, "step": 1}),
|
|
"lagrange_int": ("INT", {"default": 4, "min": 1, "max": 20, "step": 1}),
|
|
"lagrange_step": ("INT", {"default": 20, "min": 5, "max": 100, "step": 1}),
|
|
"max_interval": ("INT", {"default": 4, "min": 1, "max": 20, "step": 1}),
|
|
"acc_start": ("INT", {"default": 10, "min": 0, "max": 100, "step": 1}),
|
|
"acc_end": ("INT", {"default": 47, "min": -100, "max": 200, "step": 1}),
|
|
|
|
# === TeaCache model 参数 ===
|
|
"teacache_model_type": (TEACACHE_MODEL_TYPES, {"default": "wan2.1_t2v_1.3B_ret_mode"}),
|
|
"rel_l1_thresh": ("FLOAT", {"default": 0.4, "min": 0.0, "max": 10.0, "step": 0.01}),
|
|
"start_percent": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.01}),
|
|
"end_percent": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
|
|
"cache_device": (["cuda", "cpu"], {"default": "cuda"}),
|
|
# === MagCache model 参数 ===
|
|
"magcache_model_type": (MAGCACHE_MODEL_TYPES, {"default": "wan2.1_t2v_1.3B"}),
|
|
"magcache_thresh": ("FLOAT", {"default": 0.06, "min": 0.0, "max": 0.3, "step": 0.01}),
|
|
"retention_ratio": ("FLOAT", {"default": 0.2, "min": 0.1, "max": 0.3, "step": 0.01}),
|
|
"magcache_K": ("INT", {"default": 2, "min": 0, "max": 6, "step": 1}),
|
|
"start_step": ("INT", {"default": 0, "min": 0, "max": 100, "step": 1}),
|
|
"end_step": ("INT", {"default": -1, "min": -100, "max": 100, "step": 1}),
|
|
# === TaylorSeer model 参数 ===
|
|
"taylorseer_model_type": (TAYLORSEER_MODEL_TYPES, {"default": "flux"}),
|
|
"max_order": ("INT", {"default": 1, "min": 0, "max": 4, "step": 1}),
|
|
"fresh_threshold": ("INT", {"default": 6, "min": 2, "max": 20, "step": 1}),
|
|
"first_enhance": ("INT", {"default": 3, "min": 1, "max": 10, "step": 1}),
|
|
# === HiCache model 参数 ===
|
|
"hicache_model_type": (HICACHE_MODEL_TYPES, {"default": "flux"}),
|
|
"hicache_prediction_mode": (HICACHE_PREDICTION_MODES, {"default": "hicache"}),
|
|
"hicache_max_order": ("INT", {"default": 1, "min": 0, "max": 4, "step": 1}),
|
|
"hicache_fresh_threshold": ("INT", {"default": 6, "min": 2, "max": 20, "step": 1}),
|
|
"hicache_first_enhance": ("INT", {"default": 3, "min": 1, "max": 10, "step": 1}),
|
|
"hicache_scale_factor": ("FLOAT", {"default": 0.7, "min": 0.01, "max": 2.0, "step": 0.01}),
|
|
# === SeaCache model 参数 ===
|
|
"seacache_model_type": (SEACACHE_MODEL_TYPES, {"default": "flux"}),
|
|
"seacache_thresh": ("FLOAT", {"default": 0.3, "min": 0.0, "max": 2.0, "step": 0.01}),
|
|
"seacache_power_exp": ("FLOAT", {"default": 2.0, "min": 1.0, "max": 5.0, "step": 0.5}),
|
|
"seacache_ret_steps": ("INT", {"default": 1, "min": 0, "max": 10, "step": 1}),
|
|
# === TC-Pade model 参数 ===
|
|
"tcpade_model_type": (TCPADE_MODEL_TYPES, {"default": "wan2.1"}),
|
|
"tcpade_start_step": ("INT", {"default": 4, "min": 0, "max": 100, "step": 1}),
|
|
"tcpade_end_step": ("INT", {"default": -1, "min": -1, "max": 200, "step": 1}),
|
|
"tcpade_interval": ("INT", {"default": 8, "min": 1, "max": 100, "step": 1}),
|
|
"tcpade_n_threshold": ("FLOAT", {"default": 1.4, "min": 0.0, "max": 5.0, "step": 0.1}),
|
|
"tcpade_predictor_order": ("INT", {"default": 3, "min": 1, "max": 4, "step": 1}),
|
|
"tcpade_history_size": ("INT", {"default": 6, "min": 3, "max": 12, "step": 1}),
|
|
"tcpade_cache_device": (["cuda", "cpu"], {"default": "cuda"}),
|
|
# === ZEUS sampler 参数 (默认值来自 flux_demo.py) ===
|
|
"zeus_denominator": ("INT", {"default": 3, "min": 2, "max": 6, "step": 1}),
|
|
"zeus_modular": ("STRING", {"default": "0,1"}),
|
|
"zeus_acc_start": ("INT", {"default": 10, "min": 0, "max": 100, "step": 1}),
|
|
"zeus_acc_end": ("INT", {"default": 45, "min": -100, "max": 200, "step": 1}),
|
|
"zeus_interp_mode": (ZEUS_INTERP_MODES, {"default": "psi"}),
|
|
"zeus_caching_mode": (ZEUS_CACHING_MODES, {"default": "reuse_interp"}),
|
|
"zeus_max_interval": ("INT", {"default": 4, "min": 1, "max": 20, "step": 1}),
|
|
"zeus_lagrange_term": ("INT", {"default": 3, "min": 0, "max": 5, "step": 1}),
|
|
"zeus_lagrange_int": ("INT", {"default": 6, "min": 1, "max": 20, "step": 1}),
|
|
"zeus_lagrange_step": ("INT", {"default": 24, "min": 5, "max": 100, "step": 1}),
|
|
},
|
|
}
|
|
|
|
RETURN_TYPES = ("SAMPLER", "MODEL")
|
|
RETURN_NAMES = ("sampler", "model")
|
|
FUNCTION = "get_sampler"
|
|
CATEGORY = "AccelDiff"
|
|
|
|
@classmethod
|
|
def IS_CHANGED(cls, sampler_method, model_method, **kwargs):
|
|
return f"{sampler_method}_{model_method}"
|
|
|
|
def get_sampler(self, sampler_method="AdaptiveDiff", model_method="None",
|
|
sampler_name="euler",
|
|
model=None,
|
|
# --- AdaptiveDiff sampler 参数 ---
|
|
max_skip_steps=4,
|
|
threshold=0.01,
|
|
# --- EasyCache sampler 参数 ---
|
|
ec_threshold=0.025,
|
|
ret_steps=5,
|
|
# --- SADA sampler 参数 ---
|
|
lagrange_term=3,
|
|
lagrange_int=4,
|
|
lagrange_step=20,
|
|
max_interval=4,
|
|
acc_start=10,
|
|
acc_end=47,
|
|
# --- ZEUS sampler 参数 ---
|
|
zeus_denominator=3,
|
|
zeus_modular="0,1",
|
|
zeus_acc_start=10,
|
|
zeus_acc_end=45,
|
|
zeus_interp_mode="psi",
|
|
zeus_caching_mode="reuse_interp",
|
|
zeus_max_interval=4,
|
|
zeus_lagrange_term=3,
|
|
zeus_lagrange_int=6,
|
|
zeus_lagrange_step=24,
|
|
# --- TeaCache model 参数 ---
|
|
teacache_model_type="wan2.1_t2v_1.3B_ret_mode",
|
|
rel_l1_thresh=0.4,
|
|
start_percent=0.0,
|
|
end_percent=1.0,
|
|
cache_device="cuda",
|
|
# --- MagCache model 参数 ---
|
|
magcache_model_type="wan2.1_t2v_1.3B",
|
|
magcache_thresh=0.06,
|
|
retention_ratio=0.2,
|
|
magcache_K=2,
|
|
start_step=0,
|
|
end_step=-1,
|
|
# --- TaylorSeer model 参数 ---
|
|
taylorseer_model_type="flux",
|
|
max_order=1,
|
|
fresh_threshold=6,
|
|
first_enhance=3,
|
|
# --- HiCache model 参数 ---
|
|
hicache_model_type="flux",
|
|
hicache_prediction_mode="hicache",
|
|
hicache_max_order=1,
|
|
hicache_fresh_threshold=6,
|
|
hicache_first_enhance=3,
|
|
hicache_scale_factor=0.7,
|
|
# --- SeaCache model 参数 ---
|
|
seacache_model_type="flux",
|
|
seacache_thresh=0.3,
|
|
seacache_power_exp=2.0,
|
|
seacache_ret_steps=1,
|
|
# --- TC-Pade model 参数 ---
|
|
tcpade_model_type="wan2.1",
|
|
tcpade_start_step=4,
|
|
tcpade_end_step=-1,
|
|
tcpade_interval=8,
|
|
tcpade_n_threshold=1.4,
|
|
tcpade_predictor_order=3,
|
|
tcpade_history_size=6,
|
|
tcpade_cache_device="cuda",
|
|
**kwargs):
|
|
|
|
out_sampler = None
|
|
out_model = None
|
|
|
|
# ====== Sampler 构建 ======
|
|
if sampler_method != "None":
|
|
extra_options = {}
|
|
|
|
# --- 根据 sampler_method 导入对应的模块和 KSAMPLER ---
|
|
if sampler_method == "SADA":
|
|
from .sampler.sada import KSAMPLER as _KSAMPLER
|
|
from .sampler import sada as _sampling
|
|
_ksampler_kwargs = dict(
|
|
max_interval=max_interval,
|
|
acc_start=max(acc_start, 3),
|
|
acc_end=acc_end,
|
|
lagrange_term=lagrange_term,
|
|
lagrange_int=lagrange_int,
|
|
lagrange_step=lagrange_step,
|
|
)
|
|
elif sampler_method == "ZEUS":
|
|
from .sampler.zeus import KSAMPLER as _KSAMPLER
|
|
from .sampler import zeus as _sampling
|
|
zeus_modular_tuple = tuple(int(x.strip()) for x in zeus_modular.split(",") if x.strip())
|
|
_ksampler_kwargs = dict(
|
|
denominator=zeus_denominator,
|
|
modular=zeus_modular_tuple,
|
|
acc_start=zeus_acc_start,
|
|
acc_end=zeus_acc_end,
|
|
interp_mode=zeus_interp_mode,
|
|
caching_mode=zeus_caching_mode,
|
|
max_interval=zeus_max_interval,
|
|
lagrange_term=zeus_lagrange_term,
|
|
lagrange_int=zeus_lagrange_int,
|
|
lagrange_step=zeus_lagrange_step,
|
|
)
|
|
elif sampler_method == "AdaptiveDiff":
|
|
from .sampler.adaptivediff import KSAMPLER as _KSAMPLER
|
|
from .sampler import adaptivediff as _sampling
|
|
_ksampler_kwargs = dict(
|
|
max_skip_steps=max_skip_steps,
|
|
threshold=threshold,
|
|
)
|
|
else: # EasyCache
|
|
from .sampler.easycache import KSAMPLER as _KSAMPLER
|
|
from .sampler import easycache as _sampling
|
|
_ksampler_kwargs = dict(
|
|
threshold=ec_threshold,
|
|
ret_steps=ret_steps,
|
|
)
|
|
|
|
# --- sampler_name 映射 ---
|
|
if sampler_method == "SADA":
|
|
_sampler_name = sampler_name if sampler_name in SADA_SUPPORT_SAMPLER else "euler"
|
|
else:
|
|
_sampler_name = "euler" if sampler_name == "ddim" else sampler_name
|
|
_inpaint_opts = {"random": True} if sampler_name == "ddim" else {}
|
|
|
|
# --- dpm_fast / dpm_adaptive 需要特殊包装（签名不同） ---
|
|
if _sampler_name in ("dpm_fast", "dpm_adaptive") and sampler_method != "SADA":
|
|
_raw_func = getattr(_sampling, "sample_{}".format(_sampler_name))
|
|
_pass_keys = list(_ksampler_kwargs.keys())
|
|
|
|
def _make_wrapper(raw_func, pass_keys):
|
|
def wrapper(model, noise, sigmas, extra_args=None, callback=None, disable=None, **kw):
|
|
if len(sigmas) <= 1:
|
|
return noise
|
|
sigma_min = sigmas[-1]
|
|
if sigma_min == 0:
|
|
sigma_min = sigmas[-2]
|
|
total_steps = len(sigmas) - 1
|
|
pass_kw = {k: kw.pop(k) for k in pass_keys if k in kw}
|
|
return raw_func(model, noise, sigma_min, sigmas[0], total_steps,
|
|
extra_args=extra_args, callback=callback, disable=disable,
|
|
**pass_kw)
|
|
return wrapper
|
|
|
|
sampler_function = _make_wrapper(_raw_func, _pass_keys)
|
|
else:
|
|
sampler_function = getattr(_sampling, "sample_{}".format(_sampler_name))
|
|
|
|
# --- 统一构造 KSAMPLER ---
|
|
out_sampler = _KSAMPLER(
|
|
sampler_function,
|
|
extra_options=extra_options,
|
|
inpaint_options=_inpaint_opts,
|
|
**_ksampler_kwargs,
|
|
)
|
|
|
|
# ====== Model 构建 ======
|
|
if model_method != "None":
|
|
if model is None:
|
|
raise ValueError(f"model_method={model_method} requires a MODEL input")
|
|
|
|
if model_method == "TeaCache":
|
|
out_model = _apply_teacache(
|
|
model, teacache_model_type, rel_l1_thresh,
|
|
start_percent, end_percent, cache_device,
|
|
)
|
|
elif model_method == "MagCache":
|
|
out_model = _apply_magcache(
|
|
model, magcache_model_type, magcache_thresh,
|
|
retention_ratio, magcache_K, start_step, end_step,
|
|
)
|
|
elif model_method == "TaylorSeer":
|
|
out_model = _apply_taylorseer(
|
|
model, taylorseer_model_type, max_order,
|
|
fresh_threshold, first_enhance,
|
|
)
|
|
elif model_method == "HiCache":
|
|
out_model = _apply_hicache(
|
|
model, hicache_model_type, hicache_max_order,
|
|
hicache_fresh_threshold, hicache_first_enhance,
|
|
hicache_prediction_mode, hicache_scale_factor,
|
|
)
|
|
elif model_method == "SeaCache":
|
|
out_model = _apply_seacache(
|
|
model, seacache_model_type, seacache_thresh,
|
|
seacache_power_exp, seacache_ret_steps,
|
|
)
|
|
elif model_method == "TC-Pade":
|
|
out_model = _apply_tcpade(
|
|
model, tcpade_model_type, tcpade_start_step,
|
|
tcpade_end_step, tcpade_interval, tcpade_n_threshold,
|
|
tcpade_predictor_order, tcpade_history_size,
|
|
tcpade_cache_device,
|
|
)
|
|
else:
|
|
out_model = model
|
|
|
|
return (out_sampler, out_model)
|
|
|
|
|
|
# 节点映射
|
|
NODE_CLASS_MAPPINGS = {
|
|
"AccelDiffUnified": DynamicModelParamsNode
|
|
}
|
|
|
|
NODE_DISPLAY_NAME_MAPPINGS = {
|
|
"AccelDiffUnified": "AccelDiff Unified"
|
|
}
|