Scheduled PerpNeg CFGGuider (Inspire)

Documentation

• Class name: ScheduledPerpNegCFGGuider __Inspire
• Category: sampling/custom_sampling/guiders
• Output node: False

The ScheduledPerpNegCFGGuider node dynamically adjusts the guidance scale for negative prompts in a scheduled manner during the sampling process. It integrates negative conditioning with a configurable schedule to fine-tune the influence of negative prompts over time, aiming to optimize the generation quality by balancing the guidance scale.

Input types

Required

• model
  • Specifies the model to be used for generation, serving as the foundation for the sampling process.
  • Comfy dtype: MODEL
  • Python dtype: comfy.model_management.Model
• positive
  • Defines positive conditioning prompts that guide the generation towards desired attributes or content.
  • Comfy dtype: CONDITIONING
  • Python dtype: str
• negative
  • Specifies negative conditioning prompts to steer the generation away from undesired attributes or content.
  • Comfy dtype: CONDITIONING
  • Python dtype: str
• empty_conditioning
  • Provides an option for empty conditioning, allowing for flexibility in the generation process.
  • Comfy dtype: CONDITIONING
  • Python dtype: str
• neg_scale
  • Determines the scale of influence for negative prompts, allowing for fine-tuning of their impact on the generation.
  • Comfy dtype: FLOAT
  • Python dtype: float
• sigmas
  • Specifies the noise levels for the diffusion process, contributing to the diversity of the generated content.
  • Comfy dtype: SIGMAS
  • Python dtype: torch.Tensor
• from_cfg
  • Sets the initial guidance scale, marking the starting point of the scheduled adjustment.
  • Comfy dtype: FLOAT
  • Python dtype: float
• to_cfg
  • Defines the final guidance scale, indicating the endpoint of the scheduled adjustment.
  • Comfy dtype: FLOAT
  • Python dtype: float
• schedule
  • Determines the schedule according to which the guidance scale is adjusted over time, offering various strategies for dynamic adaptation.
  • Comfy dtype: COMBO[STRING]
  • Python dtype: str

Output types

• guider
  • Comfy dtype: GUIDER
  • Provides the configured guider with scheduled negative conditioning, ready for use in the sampling process.
  • Python dtype: Guider_PerpNeg_scheduled
• sigmas
  • Comfy dtype: SIGMAS
  • Returns the noise levels used in the diffusion process, essential for controlling the generation's randomness.
  • Python dtype: torch.Tensor

Usage tips

• Infra type: GPU
• Common nodes: unknown

Source code

class ScheduledPerpNegCFGGuider:
    @classmethod
    def INPUT_TYPES(s):
        return {"required": {
                    "model": ("MODEL", ),
                    "positive": ("CONDITIONING", ),
                    "negative": ("CONDITIONING", ),
                    "empty_conditioning": ("CONDITIONING", ),
                    "neg_scale": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01}),
                    "sigmas": ("SIGMAS", ),
                    "from_cfg": ("FLOAT", {"default": 6.5, "min": 0.0, "max": 100.0, "step": 0.1, "round": 0.01}),
                    "to_cfg": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.1, "round": 0.01}),
                    "schedule": (["linear", "log", "exp"], {'default': 'log'})
                    }
                }

    RETURN_TYPES = ("GUIDER", "SIGMAS")

    FUNCTION = "get_guider"
    CATEGORY = "sampling/custom_sampling/guiders"

    def get_guider(self, model, positive, negative, empty_conditioning, neg_scale, sigmas, from_cfg, to_cfg, schedule):
        guider = Guider_PerpNeg_scheduled(model, sigmas, from_cfg, to_cfg, schedule, neg_scale)
        guider.set_conds(positive, negative, empty_conditioning)
        return guider, sigmas