EasyUnSampler(unsampler)

EasyUnSampler runs a sampler in reverse to progressively add noise to a latent instead of denoising it. It is useful for creating noised latents for img2img-style variations, inpainting, analysis, or as the input to another sampling pass.

Description

EasyUnSampler takes the latents, positive/negative conditioning, and model from a pipeline, then walks the sampling process backward for a given number of steps. The result is a more-noised latent that can be passed to a normal sampler, allowing control over the amount of noise reintroduced.

Inputs

• steps: The number of inverse sampling steps to run. The value should usually match the step count used in the corresponding forward sampling pass.
  Default: 20, Min: 1, Max: 10000.

• end_at_step: The step at which the unsampling should stop. A value of 0 runs through all steps, producing the most-noised result. A higher value ends the process early, producing a less-noised latent.
  Default: 0, Min: 0, Max: 10000.

• cfg: The classifier-free guidance scale used during unsampling. Use 1.0 to disable guidance.
  Default: 1, Min: 0, Max: 100.

• sampler_name: The sampler to use for the inverse process. This accepts the same sampler names available elsewhere in ComfyUI.

• scheduler: The scheduler to use for the inverse process. This accepts the same scheduler names available elsewhere in ComfyUI.

• normalize: disable or enable. When enabled, the resulting latent is normalized after unsampling. This can help keep the latent values stable before another sampling pass.

• pipe: An optional PIPE_LINE input containing the model, conditioning, and latent. If given, its values are used unless overridden by the individual optional inputs below.
  Optional.

• optional_model: An optional model to use instead of the one supplied by pipe.
  Optional.

• optional_positive: An optional positive conditioning input to use instead of the one supplied by pipe.
  Optional.

• optional_negative: An optional negative conditioning input to use instead of the one supplied by pipe.
  Optional.

• optional_latent: An optional latent to begin unsampling from instead of the one supplied by pipe.
  Optional.

Outputs

• PIPE_LINE: The updated pipeline, including the unsampled latent. If a pipe was provided, it is returned with the new latent and related values. If only individual inputs were provided, a pipeline is assembled from those values.

• LATENT: The resulting unsampled latent.

Usage Notes

• You can supply all values through the pipe input, or provide optional_model, optional_positive, optional_negative, and optional_latent individually. When both are supplied, the individual optional inputs take precedence.
• end_at_step = 0 runs the entire unsampling process. To introduce less noise, set end_at_step to a step number closer to the current state of the latent.
• The output LATENT is typically connected to a normal sampler to continue generating an image from the noised latent.
• normalize is useful when the unsampled latent will be used inside another pipeline that assumes normalized latent values.