SeedVarianceEnhancer

header

https://github.com/user-attachments/assets/bdbbfa83-f8ec-4cab-81cf-36a6820a51e5

SeedVarianceEnhancer v2.2

SeedVarianceEnhancer is a ComfyUI custom node designed to add diversity to the outputs of Z-Image Turbo. It compensates for low seed variance, which is when generated images look similar despite being generated with different seeds. It works by adding random noise to the embedding for the early generation steps.

Installation:

Extract the contents of the zip file and place the SeedVarianceEnhancer directory in your ComfyUI/custom_nodes directory. Launch or restart ComfyUI. You will find SeedVarianceEnhancer inside the advanced/conditioning node group.

Usage:

The node should be placed next in line after the positive prompt "CLIP Text Encode (Prompt)" node. The conditioning output should connect to the sampler's positive prompt input. The default settings should work well with Z-Image Turbo. The node has been tested to work with the workflow from https://github.com/comfyanonymous/ComfyUI_examples/tree/master/z_image.

The randomize_percent setting determines what percentage of embedding values will be randomly selected for modification with noise. Strength adjusts the scale of the noise. The seed determines which values to modify and with what noise. Noise_insert is set to "noise on beginning steps" by default. By removing the noise before the end of the generation, the model will have time to pivot back towards prompt adherence. "Noise on all steps" will reduce prompt adherence and text rendering significantly. Steps_switchover_percent specifies what percentage of steps are processed before switching between noisy and original embeddings.

Strength values in the range of 15 to 30 and a randomize_percent of 50 are reasonable starting points for use with Z-Image Turbo. The steps_switchover_percent defaults to 20%. Optimal settings vary depending on the prompt. Some prompts will be adversely affected when using the default settings. Experimentation may be required to achieve good results. An often effective alternate configuration is to decrease steps_switchover_percent to 10 and increase strength to 40.

To ensure a specific number of steps are used before the switch, use this formula: (100/TOTALSTEPS) * STEPS - 1.

Longer more detailed prompts provide more embedding values for SeedVarianceEnhancer to manipulate. It is somewhat counterintuitive, but more detailed prompts may result in more diverse outputs.

Advanced usage:

SeedVarianceEnhancer includes an embedding masking feature, which excludes portions of the embedding from exposure to noise. Masking can protect portions of the prompt. The mask_starts_at setting controls whether the mask extends from the beginning or the end. Mask_percent controls how far the mask extends. Setting mask_starts_at to beginning and mask_percent to 50 will mask the first half of the prompt. If the prompt is structured so that critical details are in the first half, and less important details are in the second half, the image generations should more consistently adhere to the critical details.

An image can be reproduced exactly if an old workflow is reused. The prompt can be revised while maintaining a consistent output composition provided guidelines are followed. The embedding values most maintain alignment with the noise in order to produce a similar image. Appending text to the end of the prompt maintains alignment. Making single word substitutions, such as changing a color, may work but if the resulting embedding values become offset, the output image will no longer maintain consistency.

SeedVarianceEnhancer nodes can be chained together. The noise from each node will add up. Chaining does not produce more than one switchover event. The switchover schedule will follow the last node's setting.

Strength values can be set to