Sampling Methods

For image generation, Samplers are used to tell the AI how it should start generating visual patterns from the initial noise.
NovelAI supports a wide variety of Sampling Methods:
DPM++ 2M, Euler Ancestral, Euler, DPM2, DPM++ 2S Ancestral, DPM++ SDE, DPM Fast, and DDIM.

Different Samplers generate different images, but the differences may not always be obvious or predictable.

Sampler comparison.

Goose Tip: It's easier to see the differences between Samplers on images with very low Steps values. But such low values aren't recommended for normal image generation.

DPM++ 2M and Euler_Ancestral are our recommended samplers, due to their consistent, high-quality generations in combination with NovelAI Diffusion. We highly recommend leaving the Sampler setting as is, unless you have a deeper knowledge of Image Generation.

Special Samplers: SMEA & SMEA DYN

What is NAI SMEA?

Sinusoidal Multipass Euler Ancestral (SMEA) is a new sampler developed with the goal of improving overall coherency and quality, especially at higher resolutions. Based on the Euler ancestral sampler, we created a new sine-based schedule that interpolates between multiple passes of the regular diffusion model as sampling progresses. This approach ensures that Stable Diffusion attends to both local and global features. Oftentimes when sampling at higher resolutions with the conventional samplers, Stable Diffusion can produce repeats of the same subjects or bizarre anatomy, this is largely due to poor global attention at higher resolutions NAI SMEA aims to solve that.

All the normal samplers have SMEA versions and SMEA DYN variants.
(Except for the DDIM sampler, which doesn't support SMEA.)

SMEA & SMEA DYN are perfect for higher resolutions!

The differences between these two samplers variations are most obvious for higher resolution images. While both work better than normal samplers for high resolution image generation, SMEA DYN focuses less time on lower generations and begins to shine dynamically in the mid to high range of a generation. We've noticed slightly more refined compositions.
Ultimately, the effectiveness of each sampler will depend on the specific use case and image resolution.


Why is SMEA more expensive?
SMEA samplers run multiple passes of the Unet during each step. It costs slightly more than the normal samplers based on the increased compute cost, the increase is relative to the increased amount of time it takes the GPU to process a request.

Auto SMEA Toggle Option

Auto nai_smea toggle

There is a default option to Auto apply SMEA for images over 1024x1024 and above pixels, so you don't have to worry about when you should or shouldn't use it.

Goose Tip: With the Auto setting enabled, you can still easily swap between using just SMEA or SMEA DYN. You can also disable the Auto toggle and generate high resolution images without SMEA, but it's not recommended!


SMEA and SMEA DYN samplers react differently to the Step and Prompt Guidance settings. For example, the Guidance values can go much higher before becoming unpredictable. You may want to increase the Guidance value if you are generating at higher resolutions.

If you have a particularly short prompt that seemingly performs badly, you can try copy-pasting it into the prompt field repeatedly. We've noted that this can solve issues during testing.

Sometimes the sampler may focus on different aspects of your prompt that differ from the previous defaults:
You can alleviate this with emphasis and de-emphasizing aspects of your prompt.
(See Strengthening & Weakening Vectors)

We recommend experimenting to find your preferred values with samplers!

Sampler Comparisons

These high resolution images were all generated using the same prompt, settings and seed. The only differences between them were the Samplers as well as the SMEA and DYN toggles.

NAI SMEA comparison.