What is Sampling?

The image generation process begins by creating a random image in the latent space. Subsequently, a noise predictor assesses the noise present in the image, which is then subtracted. This iterative process is repeated multiple times until a clean image is obtained.

This iterative denoising procedure is called sampling because each step involves generating a new sample image. The specific method employed in this sampling process is referred to as the sampler or sampling method.

Available Samplers

The Euler Sampler

Euler, in essence, represents the simplest form of a sampler. It mirrors Euler's method, a mathematical technique utilized for solving ordinary differential equations. Notably, it operates in a fully deterministic manner, devoid of any addition of random noise during the sampling process.

Or to put it simply, the sampler is your best option if you are trying to generate image variations that won’t look too different from one another.

The Euler Ancestral Sampler

The Euler Ancestral sampler shares similarities with Euler's sampler. However, ancestral samplers diverge in their approach by subtracting excessive noise at each step and reintroducing random noise to align with the designated noise schedule. As a consequence, the denoised image is influenced by the particular noise introduced in previous steps. Therefore, the denoising path of the image is contingent upon the unique random noises added at each stage. Consequently, repeating the process would yield different results.

So in short, the sampler is able to provide a great variety of unique images based on the produced noise on each generation step.

The DPM++ 2M Karras Sampler

The DMP++ 2M Karras is an advanced variant of the original DMP++ sampler, designed by Karras, boasting enhanced capabilities and efficiency. With a focus on precision and scalability, this iteration optimizes the Denoising Diffusion Probabilistic Model (DMP++) by incorporating Karras' innovative modifications. Its key feature lies in its ability to handle high-resolution images, leveraging a sophisticated algorithm to ensure optimal denoising outcomes. By harnessing the power of cutting-edge technology, DMP++ 2M Karras delivers superior performance, making it a preferred choice for demanding denoising tasks.

To put it simply, the DPM++2M Karras Sampler is good at generating images with great detail, with vibrant color and with much less noise.

Differences Between Samplers

In terms of generated image quality, Euler and Euler Ancestral (Euler a) present notable differences. Euler, following a deterministic approach, produces images without introducing any random noise during sampling. As a result, its denoised images may exhibit a more predictable outcome with a consistent level of noise reduction. Conversely, Euler a's strategy of subtracting additional noise at each step and reintroducing random noise to match a specified schedule introduces a level of unpredictability into the denoising process. While this technique allows for potential refinement in image quality, it may also lead to varied outcomes, influenced by the specific random noises incorporated at each stage.

Comparatively, DMP++ 2M Karras sets itself apart by prioritizing denoising efficiency and scalability over deterministic noise manipulation. With a focus on handling high-resolution images, this variant of the DMP++ sampler leverages advanced algorithms to achieve superior image quality. By incorporating innovative modifications by Karras, DMP++ 2M Karras optimizes the denoising process, resulting in images with enhanced clarity and reduced noise levels. This emphasis on precision and scalability makes DMP++ 2M Karras a preferred choice for tasks requiring high-fidelity image reconstruction.

Examples of Different Samplers

So how about we explore how it really feels to generate the same image using different samplers?

Let’s use the following prompt and negative prompt:

cinematic photography of a ((beautiful blond woman wearing a red dress), ( beautiful detailed eyes), (on a balcony), (light reflections), (intricate:0.87), (anatomically correct, raw camera focus, hdr, life like shadows and highlights and reflections, natural colors, ultra detailed, soft focus, 8k, best quality, best image)

(low resolution, low detail, bad quality, bad image, deformed, extra fingers, oversaturated, ((illustration, design, artwork, painting, sketch, anime, cartoon, 3d render)), wrong orientation, out of proportion, clashing elements, ((jpeg artifacts, signature, watermark, text, letters))

This is how our image looks like with each sampler:

Result
euler.webp

Ready to Give it a Try?

Take action now by implementing these steps and creating your own images on sexy.ai!