Update SMEA Dy CFG++ samplers

README.md

@@ -1,7 +1,7 @@
 # ComfyUI-ppm
-Just a bunch of random nodes modified/fixed/created by me or others. If any node starts throwing errors after an update - try re-adding it.
+Just a bunch of random nodes modified/fixed/created by me or others. If any node starts throwing errors after an update - try to delete and re-add the node.
 
-I'll add some example workflows soon.
+I'll probably add some example workflows in the future (but I'm kinda lazy, kek).
 
 # Nodes
 
@@ -11,22 +11,25 @@ Modified implementation of NegPiP by [laksjdjf](https://github.com/laksjdjf) and
 
 `CLIPNegPip` node allows you to use negative weights in prompts. You should connect CLIPNegPip before other model/clip patches. After that, you can enter negative weights in your prompts (CTRL + arrows hotkey is capped at 0.0, will probably fix that soon).
 
-You can read more about NegPiP [in the original repo](https://github.com/hako-mikan/sd-webui-negpip). I recommend putting everything from negative prompt to positive with a negative weight of something like -1.1 or -1.3. It's also better to keep all commas inside weight braces (i.e. `(worst quality,:-1.3) (sketch:-1.1,)` instead of `(worst quality:-1.3), (sketch:-1.1),`).
+Read more about NegPiP [in the original repo](https://github.com/hako-mikan/sd-webui-negpip). I recommend putting everything from negative prompt to positive with a negative weight of something like -1.1 or -1.3. It's also better to keep all commas inside weight braces (i.e. `(worst quality,:-1.3) (sketch:-1.1,)` instead of `(worst quality:-1.3), (sketch:-1.1),`).
 
 ## AttentionCouplePPM
 Modified implementation of AttentionCouple by [laksjdjf](https://github.com/laksjdjf) and [Haoming02](https://github.com/Haoming02). I made `AttentionCouplePPM` node compatible with `CLIPNegPiP` node and with default `PatchModelAddDownscale (Kohya Deep Shrink)` node. You can add/remove regions by right-clicking the node and selecting `Add Region`/`Remove Region`.
 
-You can use multiple `LatentToMaskBB` nodes to set bounding box masks for `AttentionCouplePPM`. The parameters are relative to your initial latent: `x=0.5, y=0.0, w=0.5, h=1.0` will produce a mask covering right half of your image.
+Use multiple `LatentToMaskBB` nodes to set bounding box masks for `AttentionCouplePPM`. The parameters are relative to your initial latent: `x=0.5, y=0.0, w=0.5, h=1.0` will produce a mask covering right half of your image.
 
 ## CFG++SamplerSelect
-Samplers adapted to [CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models by Chung et al.](https://cfgpp-diffusion.github.io/)
+Samplers adapted to [CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models by Chung et al.](https://cfgpp-diffusion.github.io/).
+Includes some samplers from [Euler-Smea-Dyn-Sampler by Koishi-Star](https://github.com/Koishi-Star/Euler-Smea-Dyn-Sampler).
 
-Should greatly reduce overexposure effect. Use together with `SamplerCustom` node. Don't forget to set CFG scale to 1.0-2.0 and PAG scale (if used) to 0.5-1.0.
+Should greatly reduce overexposure effect. Use together with `SamplerCustom` node. Don't forget to set CFG scale to 1.0-2.0 and PAG/SEG scale (if used) to 0.5-1.0.
+
+Tweak `s_dy_pow` parameter while using `*_dy_*` samplers to reduce blur artifacts (value `-1` disables this feature).
 
 ## Guidance Limiter
 Implementation of [Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models by Kynkäänniemi et al.](https://arxiv.org/abs/2404.07724) (also contains `RescaleCFG` functionality)
 
-Guidance Limiter is also available in a format of guider node `CFGLimiterGuider` for `SamplerCustomAdvanced`.
+Guidance Limiter is also available as a `CFGLimiterGuider` guider node for `SamplerCustomAdvanced`.
 
 ## Empty Latent Image (Aspect Ratio)
 `Empty Latent Image (Aspect Ratio)` node generates empty latent with specified aspect ratio and with respect to target resolution.

ppm_cfgpp_dyn_sampling.py

@@ -5,6 +5,9 @@
 from comfy.k_diffusion import sampling
 import comfy.model_patcher
 
+CFGPP_SAMPLER_NAMES_DYN = ["euler_dy_cfg_pp", "euler_smea_dy_cfg_pp"]
+CFGPP_SAMPLER_NAMES_DYN_ETA = []
+
 
 class _Rescaler:
  def __init__(self, model, x, mode, **extra_args):
@@ -81,7 +84,7 @@ def post_cfg_function(args):
 
 @torch.no_grad()
 def sample_euler_dy_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0.,
- s_tmax=float('inf'), s_noise=1., s_gamma=None):
+ s_tmax=float('inf'), s_noise=1., s_dy_pow=-1):
  extra_args = {} if extra_args is None else extra_args
  s_in = x.new_ones([x.shape[0]])
 
@@ -95,8 +98,8 @@ def post_cfg_function(args):
 
  for i in trange(len(sigmas) - 1, disable=disable):
  gamma = max(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
- if s_gamma is not None:
- gamma = s_gamma
+ if s_dy_pow >= 0:
+ gamma = gamma * (1.0 - (i / (len(sigmas) - 2))**s_dy_pow)
  sigma_hat = sigmas[i] * (gamma + 1)
  # print(sigma_hat)
  dt = sigmas[i + 1] - sigma_hat
@@ -137,7 +140,7 @@ def post_cfg_function(args):
 
 @torch.no_grad()
 def sample_euler_smea_dy_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0.,
- s_tmax=float('inf'), s_noise=1.):
+ s_tmax=float('inf'), s_noise=1., s_dy_pow=-1):
  extra_args = {} if extra_args is None else extra_args
  s_in = x.new_ones([x.shape[0]])
 
@@ -151,6 +154,8 @@ def post_cfg_function(args):
 
  for i in trange(len(sigmas) - 1, disable=disable):
  gamma = max(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
+ if s_dy_pow >= 0:
+ gamma = gamma * (1.0 - (i / (len(sigmas) - 2))**s_dy_pow)
  sigma_hat = sigmas[i] * (gamma + 1)
  dt = sigmas[i + 1] - sigma_hat
  if gamma > 0:

ppm_cfgpp_sampling.py

@@ -4,6 +4,21 @@
 import comfy.model_patcher
 from comfy.k_diffusion.sampling import BrownianTreeNoiseSampler
 
+CFGPP_SAMPLER_NAMES_KD = [
+ "dpmpp_2m_cfg_pp",
+ "dpmpp_2m_sde_cfg_pp",
+ "dpmpp_2m_sde_gpu_cfg_pp",
+ "dpmpp_3m_sde_cfg_pp",
+ "dpmpp_3m_sde_gpu_cfg_pp",
+]
+
+CFGPP_SAMPLER_NAMES_KD_ETA = [
+ "dpmpp_2m_sde_cfg_pp",
+ "dpmpp_2m_sde_gpu_cfg_pp",
+ "dpmpp_3m_sde_cfg_pp",
+ "dpmpp_3m_sde_gpu_cfg_pp",
+]
+
 
 @torch.no_grad()
 def sample_dpmpp_2m_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "comfyui-ppm"
 description = "Fixed AttentionCouple/NegPip(negative weights in prompts), more CFG++ samplers, etc."
-version = "1.0.6"
+version = "1.0.7"
 license = { text = "GNU Affero General Public License v3" }
 
 [project.urls]

samplers.py

@@ -5,14 +5,19 @@
 
 INITIALIZED = False
 CFGPP_SAMPLER_NAMES_ORIGINAL = ["euler_cfg_pp", "euler_ancestral_cfg_pp"]
-CFGPP_SAMPLER_NAMES_DYN = ["euler_dy_cfg_pp", "euler_smea_dy_cfg_pp"]
-CFGPP_SAMPLER_NAMES = CFGPP_SAMPLER_NAMES_ORIGINAL + [
- "dpmpp_2m_cfg_pp",
- "dpmpp_2m_sde_cfg_pp",
- "dpmpp_2m_sde_gpu_cfg_pp",
- "dpmpp_3m_sde_cfg_pp",
- "dpmpp_3m_sde_gpu_cfg_pp",
-] + CFGPP_SAMPLER_NAMES_DYN
+CFGPP_SAMPLER_NAMES_ORIGINAL_ETA = ["euler_ancestral_cfg_pp"]
+
+
+CFGPP_SAMPLER_NAMES = [
+ *CFGPP_SAMPLER_NAMES_ORIGINAL,
+ *ppm_cfgpp_sampling.CFGPP_SAMPLER_NAMES_KD,
+ *ppm_cfgpp_dyn_sampling.CFGPP_SAMPLER_NAMES_DYN,
+]
+CFGPP_SAMPLER_NAMES_ETA = [
+ *CFGPP_SAMPLER_NAMES_ORIGINAL_ETA,
+ *ppm_cfgpp_sampling.CFGPP_SAMPLER_NAMES_KD_ETA,
+ *ppm_cfgpp_dyn_sampling.CFGPP_SAMPLER_NAMES_DYN_ETA,
+]
 
 
 def inject_samplers():
@@ -29,6 +34,7 @@ def INPUT_TYPES(s):
  "required": {
  "sampler_name": (CFGPP_SAMPLER_NAMES,),
  "eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01, "round": False}),
+ "s_dy_pow": ("INT", {"default": -1, "min": -1, "max": 100}),
  }
  }
 
@@ -37,13 +43,17 @@ def INPUT_TYPES(s):
 
  FUNCTION = "get_sampler"
 
- def get_sampler(self, sampler_name, eta: float):
+ def get_sampler(self, sampler_name, eta: float, s_dy_pow: int):
  if sampler_name in CFGPP_SAMPLER_NAMES_ORIGINAL:
  sampler_func = getattr(k_diffusion_sampling, "sample_{}".format(sampler_name))
- elif sampler_name in CFGPP_SAMPLER_NAMES_DYN:
- sampler_func = getattr(ppm_cfgpp_dyn_sampling, "sample_{}".format(sampler_name))
- else:
+ elif sampler_name in ppm_cfgpp_sampling.CFGPP_SAMPLER_NAMES_KD:
  sampler_func = getattr(ppm_cfgpp_sampling, "sample_{}".format(sampler_name))
- extra_options = {} if sampler_name in {"euler_cfg_pp", "dpmpp_2m_cfg_pp", "euler_dy_cfg_pp", "euler_smea_dy_cfg_pp"} else {"eta": eta}
+ elif sampler_name in ppm_cfgpp_dyn_sampling.CFGPP_SAMPLER_NAMES_DYN:
+ sampler_func = getattr(ppm_cfgpp_dyn_sampling, "sample_{}".format(sampler_name))
+ extra_options = {}
+ if sampler_name in CFGPP_SAMPLER_NAMES_ETA:
+ extra_options["eta"] = eta
+ if sampler_name in ppm_cfgpp_dyn_sampling.CFGPP_SAMPLER_NAMES_DYN:
+ extra_options["s_dy_pow"] = s_dy_pow
  sampler = KSAMPLER(sampler_func, extra_options=extra_options)
  return (sampler,)