In the context of k-diffusion samplers, Eta (η) is a parameter that controls the amount of randomness (stochasticity) during the sampling process. It influences how noisy or deterministic the sampling trajectory is when generating an image.
When it comes to flaws, this can help a great deal in correcting those.
Understanding Eta (η) in k-diffusion Samplers:
- η = 0 (Deterministic Sampling)
- The sampling process follows a purely deterministic path.
- This results in a more stable and repeatable generation of images, meaning the same input (latent noise + seed) will always produce the same output.
- This is typically used in ancestral samplers when you want precise control over the result.
- η > 0 (Stochastic Sampling)
- When η > 0, additional noise is introduced during the reverse diffusion steps.
- This increases variability, meaning the same input could generate different outputs.
- A small η leads to slight variations, while a larger η results in more diversity but potentially more chaotic results.
- This is useful for exploring different variations of an image while still following the general guidance of the prompt.
- η = 1 (Maximum Stochasticity in Ancestral Samplers)
- This allows for more diverse and creative outputs, useful when generating multiple variations of an image.
How η Works in Different k-Diffusion Samplers
- In samplers like Euler a (Euler ancestral) or DPM++ 2M Karras, η is an important factor in balancing determinism and randomness.
- Ancestral samplers (Euler a, DPM2 a, etc.) use η to add noise at each step.
- Non-ancestral samplers (Euler, DPM2, etc.) typically do not use η or set it to zero by default.
When to Use Different η Values
- For reproducibility: Use η = 0 (fully deterministic).
- For more creative variations: Use a small η (0.1–0.3).
- For high diversity and randomness: Use η close to 1.
Would you like help tuning η for a specific k-diffusion sampler in your workflow?
