Over the last couple of years, I've noticed a lot of ambiguity in articles and posts related to "which sampler to use and when." I've seen a number of really well-intentioned articles and posts (not just here) that immediately rely on grids and image comparisons, but those really don't do as good of a job of explaining what's going on as the authors would like. The challenge in showing a comparison is that tastes, models, and subject matter all impact "what's best."
Anyway, I thought I'd sit down and just write what each sampler/scheduler does and how they might be best combined to achieve what you're looking to accomplish.
So, first, I'll go through each sampler and scheduler and explain how they impact things. Then, I'll show you some combos and why they'd be a good choice for SD1.5 or SDXL/Pony and low-step or high-step generations. I don't have enough on Flux yet to include it, but I'm sure I'll update this later.
SUPER IMPORTANT—if I misspeak, let me know. I work with AI daily, but I'm at the executive level and not in the trenches with my engineers. SO... I might say something wrong (but this is all written as I understand it).
What Are Samplers and Why Do They Matter?
In Stable Diffusion, samplers guide the process of turning noise into an image over multiple steps. The sampler controls the diffusion process—how each image layer is iteratively improved, transitioning from a random noise field to something recognizable and detailed. The number of steps you use (high or low) and the combination of samplers and schedulers you choose will determine your generated images' style, quality, and speed.
Some samplers focus on speed, like Euler A, which is excellent when you need a quick result, while others, like DDIM or DPM++, focus more on precision and control, making them ideal for high-detail work but often taking more time.
Here are the samplers that are present in my current version of Forge:
DPM++ Family
These samplers use a diffusion probabilistic model with improvements aimed at achieving high-quality results while remaining computationally efficient. They are known for producing sharper and more consistent images, especially for realistic styles.
DPM++ 2M: A deterministic sampler that aims for precise, smooth image transitions. It's suitable for realism but may take longer than others.
Strength: High precision.
Limitation: Slightly slower than simpler samplers.
DPM++ SDE: Stochastic differential equations (SDE) introduce randomness into the sampling process. This variant helps explore more diverse outputs, which can improve detail in complex or textured areas of the image.
Strength: Enhances diversity.
Limitation: Results can vary slightly between runs.
DPM++ 2M SDE: Combines the deterministic (2M) approach with SDE. Offers the control of deterministic methods with the flexibility of SDE, making it versatile.
Strength: Balanced diversity and control.
Limitation: It may not always be the fastest option.
DPM++ 2M SDE Heun: Uses the Heun method (a specific numerical integration technique) to improve image transitions, reducing noise. This works well for more photorealistic outputs.
Strength: Clean, smooth outputs.
Limitation: Longer time due to integration steps.
DPM++ 2S a: Another variant with different weighting strategies. It can sometimes give more stylized or artistic results.
Strength: Good for artistic images.
Limitation: It may not always be realistic.
DPM++ 3M SDE: A higher-order 2M SDE version capable of producing even more intricate details but might be slower.
Strength: Great for highly detailed images.
Limitation: Computationally heavy.
Euler Family
Euler samplers are known for their speed and effectiveness. They work well with lower steps and tend to be more efficient computationally.
Euler A: A modified Euler method is known for speed and producing bold, consistent results. It's often used for fast generation but might miss out on fine details.
Strength: Speed and consistency.
Limitation: Lack of fine detail in lower step counts.
Euler: The standard Euler method performs better on detail than Euler A but can be less efficient in terms of time.
Strength: Good for balanced images.
Limitation: May struggle with very complex details.
LMS (Laplacian Pyramid Sampling)
This is a popular sampler for realistic images, especially when paired with higher step counts. It offers robust control over image transitions and produces results that often appear clean and natural.
Strength: Great for photorealism and high-quality detail.
Limitation: Slower compared to faster samplers like Euler A.
Heun
Heun samplers use an adaptive step mechanism that tries to improve transition smoothness. They are slightly slower but often result in smoother, less noisy images.
Strength: Smooth transitions and noise reduction.
Limitation: Time-consuming for larger images.
DPM2 / DPM2 A
These are second-order samplers designed to be fast while maintaining reasonable detail and realism.
DPM2: The base model is good for general use with a balance of speed and quality.
Strength: Fast and good quality.
Limitation: Sometimes misses out on intricate details.
DPM2 a: Adds an "a" variant to increase diversity, producing slightly more varied results than DPM2.
Strength: Good diversity.
Limitation: A bit slower than DPM2.
DPM Fast / DPM Adaptive
These are optimized samplers designed for speed with adaptive techniques to adjust the sampling process dynamically.
DPM Fast: Prioritizes speed over detail, making it suitable for quick iterations.
Strength: Extremely fast.
Limitation: Loss of fine detail.
DPM Adaptive: Adjusts step size dynamically, balancing speed and detail depending on the complexity of the image.
Strength: Great balance for complex tasks.
Limitation: Results can vary with different complexities.
Other Specialized Samplers
Restart: A method that periodically resets parts of the image generation to avoid getting stuck in local optima, leading to more varied results.
Strength: Adds randomness and creativity.
Limitation: This can sometimes lead to less coherent results.
HeunPP2: A variant using more advanced numerical methods for improved control over fine details.
Strength: High precision.
Limitation: Slower than Heun.
IPNDM / IPNDM_V: Iterative samplers designed for higher precision with multi-step processes, especially useful for generating realistic details.
Strength: Very high-quality outputs.
Limitation: Time-intensive.
DEIS: Diffusion-inspired sampler designed to balance speed and quality.
Strength: Versatile with good realism.
Limitation: Not as fast as Euler A.
DDIM (Denoising Diffusion Implicit Models)
This sampler is highly regarded for its ability to create detailed, realistic images while being relatively fast. DDIM works well with both lower and higher steps.
Strength: Good realism with efficient sampling.
Limitation: It miss some finer details compared to samplers like LMS.
PLMS (Pseudo-Laplacian Pyramid Sampling)
Another popular method for realism, but it's a bit faster than LMS. It works well for balanced, photorealistic outputs.
Strength: Great for high-quality results without too much time.
Limitation: Slightly less control than LMS.
UniPC
A newer sampler that focuses on unifying the strengths of previous samplers while being efficient and high quality. It's versatile for different image styles.
Strength: Versatile and efficient.
Limitation: New, so it might have less proven consistency.
LCM (Latent Control Models)
This sampler introduces latent space control, giving more refined control over how features emerge in the final image.
Strength: High control for specific details.
Limitation: Complexity and time.
DDPM (Denoising Diffusion Probabilistic Models)
The original diffusion model is known for being the most stable and slowest. It's excellent for high-quality images but not ideal if speed is needed.
Strength: Extremely stable and consistent.
Limitation: Slowest of the bunch.
Specialized Euler Variants
Euler_Dy_Negative / Euler_Dy: These are dynamic Euler samplers optimized for different image styles, beneficial for more artistic or experimental tasks.
Strength: Creative results.
Limitation: Less photorealism.
Euler_Max / Euler_Negative: More extreme versions of Euler that focus on particular dynamics of noise and structure.
Strength: Bold images.
Limitation: Not always realistic.
Euler_Smea_Dy / Euler_Smea: Smear-based methods that give images a slightly softened, dream-like quality.
Strength: Good for softer, more artistic images.
Limitation: Less sharp.
When choosing which samplers to use with SD1.5 models vs. SDXL models, several factors impact the performance, quality, and speed of the image generation process. Both versions of Stable Diffusion have distinct architectures, with SDXL being significantly larger and more complex than SD1.5, and these differences can affect how well specific samplers perform. Here's an overview of the advantages and disadvantages of using various samplers with each model:
Model Complexity and Size
Stable Diffusion 1.5: A smaller, more compact model that works well with various samplers. It's known for faster generation times, even on lower-end hardware, and it's relatively easy to control with simpler samplers like Euler A.
SDXL: A larger model designed to improve realism and detail, which inherently benefits from more advanced samplers but can be slower and require more computational resources. It's built to leverage the strengths of higher-order samplers.
Advantages/Disadvantages with Samplers
DPM++ Family
SD1.5: DPM++ samplers are highly effective with SD 1.5 models, producing sharp, clean results while keeping computational load manageable. However, for lower-resolution outputs, the extra complexity may not be necessary.
Advantage: DPM++ 2M, 2S, and other variants offer a solid balance of quality and speed.
Disadvantage: Overkill for simple tasks or low-step counts.
SDXL: These samplers shine with SDXL due to the model's larger size and capacity for more intricate detail. SDXL can leverage the extra refinement that DPM++ samplers provide, especially in photorealistic images.
Advantage: Excellent control over finer details and textures, particularly with DPM++ 2M SDE Heun and similar variants.
Disadvantage: This can be slower due to SDXL's larger architecture, especially on lower-end hardware.
Euler A / Euler
SD1.5: Euler A is often a go-to for SD 1.5 because of its speed and efficiency. It's excellent for quick previews or more straightforward compositions, though it might not produce the finest details.
Advantage: Fast, efficient, good enough for most use cases.
Disadvantage: Lacks some fine detail and precision for high-quality images.
SDXL: Euler samplers are generally less effective for SDXL because they don't handle the larger model's complexity as well. The result is often less sharp or too simplistic than other samplers designed for higher fidelity.
Advantage: Speed can be a benefit if you're running low steps.
Disadvantage: Misses out on the finer details SDXL is designed to produce.
LMS
SD1.5: LMS is well-suited to SD 1.5 for generating realistic images, especially when you want clean transitions and a good balance between speed and quality.
Advantage: Excellent for photorealism without excessive computational load.
Disadvantage: Slightly slower than Euler A for previews.
SDXL: LMS works well with SDXL, especially for high-resolution, detailed outputs. The combination of LMS and SDXL can create lifelike images, but the slower speed might be more noticeable due to SDXL's size.
Advantage: Great balance for realism in SDXL.
Disadvantage: Can be slow, especially on lower-end GPUs.
Heun / HeunPP2
SD1.5: Heun samplers are solid performers, with SD 1.5 for images requiring smooth transitions and less noise. They aren't the fastest, but they produce consistent results.
Advantage: Smooth, noise-free images.
Disadvantage: Slower compared to simpler samplers like Euler A.
SDXL: Heun variants perform very well with SDXL, especially HeunPP2 because they handle the model's need for smoothness and detail across large image sizes. However, the larger the model, the more significant the speed tradeoff.
Advantage: Ideal for smooth photorealistic transitions.
Disadvantage: Much slower than on SD 1.5 due to SDXL's size.
DPM Fast / DPM Adaptive
SD1.5: These samplers are great when you need a fast result but don't provide the highest detail. They work best for low-step counts or quick iteration in SD 1.5.
Advantage: High speed for previews or simple images.
Disadvantage: Lack of detail in complex or large images.
SDXL: DPM Fast and Adaptive lose much of their effectiveness with SDXL, as SDXL's complexity demands more refined sampling to achieve its full potential. You may still use them for quick drafts, but the final image quality may suffer.
Advantage: Speed but limited usefulness.
Disadvantage: Fails to take full advantage of SDXL's capabilities.
DDIM
SD1.5: DDIM works well for lower and higher step counts in SD 1.5, providing a good middle ground between speed and quality.
Advantage: Consistent, high-quality results without a considerable time tradeoff.
Disadvantage: Sometimes not as sharp as DPM++ on very detailed images.
SDXL: DDIM continues to be effective with SDXL. It offers excellent realism and benefits from SDXL's more extensive data. While slightly slower than with SD 1.5, the results are worth it for larger, more complex images.
Advantage: Balanced performance for SDXL's complexity.
Disadvantage: Slower than on SD 1.5, but not excessively so.
PLMS
SD1.5: PLMS is excellent for detailed, realistic images in SD 1.5, producing cleaner images with less noise than some other samplers.
Advantage: High-quality output with relatively quick convergence.
Disadvantage: Not the fastest for previews or quick generation.
SDXL: PLMS shines with SDXL, particularly when generating photorealistic, highly detailed images. It helps to avoid noise and artifacts, making it a great choice for more complex compositions.
Advantage: Great for high-end photorealism with SDXL.
Disadvantage: Slower than on SD 1.5, but highly effective.
Other Samplers for SDXL
UniPC: This sampler is particularly suited to SDXL's larger architecture. It helps produce efficient, high-quality images without excessive computational load.
Advantage: Great for SDXL's complex structure.
Disadvantage: Newer, so not as widely tested for consistency.
DEIS: Works better with SDXL than SD1.5 because of its ability to leverage larger models for high-quality outputs while maintaining efficiency.
Advantage: Good for balancing speed and quality in SDXL.
Disadvantage: SD1.5 doesn't see as much of a benefit from DEIS.
Advantages of Samplers in SDXL vs. SD1.5
SDXL Benefits More from Advanced Samplers: SDXL has more detail to leverage from higher-order samplers like DPM++, PLMS, and Heun, which handle complexity and texture better.
SD1.5 Can Use Simpler Samplers: Simpler samplers like Euler A or DPM Fast work well with SD 1.5 because the model is smaller and faster to run. You can get away with faster samplers and still achieve decent results.
Realism and Complexity: If realism and photorealistic textures are your priority, SDXL paired with LMS, PLMS, or DPM++ Heun will produce much better results than SD 1.5.
Disadvantages of Samplers in SDXL vs. SD1.5
Speed Tradeoffs in SDXL: Larger models like SDXL are inherently slower. Even though advanced samplers produce better results, they can significantly slow down the process, especially on lower-end hardware.
Overkill for Simple Tasks: Using high-order samplers like DPM++ 3M or HeunPP2 with SD 1.5 might be overkill unless you need high-quality details. SD 1.5 can produce great results with simpler, faster samplers.
For SDXL, advanced samplers like DPM++, Heun, PLMS, and LMS will bring out the model's full potential, producing more detailed, realistic images at the cost of speed. With SD 1.5, simpler samplers like Euler a, DDIM, or DPM Fast are sufficient for most tasks and offer faster generation times, making them better for quick previews or simpler compositions.
The Role of Schedulers
Schedulers work alongside samplers by controlling how noise is reduced at each step of the diffusion process. Think of them as the guides for when and how aggressively to denoise the image. The right scheduler can make a massive difference in quality, especially at lower step counts. For example, with its logarithmic curve, the Karras scheduler spreads the noise reduction more evenly, allowing for better results at lower step counts without sacrificing detail.
Scheduler Settings
Use Same Scheduler: Ensures that the same scheduler is applied consistently throughout the generation process.
SD1.5: This is generally good for stability and repeatability. Consistent scheduler usage can produce more predictable results, mainly using faster samplers like Euler A.
SDXL: Useful in SDXL to maintain control over more complex outputs. Since SDXL is highly detailed, this setting can help preserve the details that emerge from the large model.
Automatic: Let the system automatically pick the scheduler based on the sampler and other parameters.
SD1.5: Handy for general purposes and quick experiments, especially when unsure which scheduler to use. The system usually selects a suitable option.
SDXL: It might be useful for quick prototyping or exploring different styles. However, it may not always pick the best combination for maximum quality.
Uniform: In uniform scheduling, each step receives the same amount of noise reduction. It is linear and consistent.
SD1.5: Good for simpler, faster tasks or when you don't need a complex noise decay process. Results are often decent but can lack detail in complex areas.
SDXL: It's less effective in SDXL, where more advanced schedulers tend to bring out the best in the model's complexity. Uniform scheduling might oversimplify the process, reducing image quality.
Karras: The Karras scheduler uses a logarithmic noise schedule, which tends to distribute more refinement toward the later generation stages. It's designed to improve high-frequency detail without sacrificing speed.
SD1.5: This scheduler is excellent when you want to achieve detailed outputs without too many steps. It works incredibly well with DDIM or LMS.
SDXL: It excels in SDXL for generating realistic and highly detailed images. Since SDXL benefits from fine detail, Karras helps to sharpen those details later in the process without bogging down early steps.
Exponential: In exponential scheduling, noise reduction follows an exponential curve, leading to rapid refinement at the start and slower transitions toward the end.
SD1.5: Useful for quick, rough drafts. Early refinement can help you get a basic structure fast, but the results can be less detailed.
SDXL: Less ideal for SDXL because the model benefits from detailed transitions, especially in the later stages. It may cause the output to look too rough or unfinished.
Polyexponential: A variant of exponential scheduling, but with more control over the polynomial decay of noise. This allows for finer tuning.
SD1.5: Great for balancing speed and refinement. It can produce higher-quality images than standard exponential without slowing down too much.
SDXL: This can be more useful in SDXL than the basic exponential schedule, as it allows more control over the refinement of details but still doesn't compete with more advanced schedulers like Karras.
SGM Uniform: SGM (Score-based Generative Models) use a uniform noise schedule, which tends to balance noise reduction across steps but may favor smoother transitions.
SD1.5: Good for producing consistent results, though it's not the most detail-oriented scheduler. Works well for smoother, less detailed images.
SDXL: It's less effective in SDXL for high-detail images but can work well for smoother, softer images.
KL Optimal: KL (Kullback-Leibler) optimal schedulers minimize the KL divergence between the target and current distribution, resulting in a more mathematically "optimal" diffusion process.
SD1.5: Good for producing images with consistent structure and balance. Works well when combined with DDIM or LMS.
SDXL: Works well for complex images in SDXL because it focuses on preserving details in a mathematically optimal way. It can help generate realistic, well-structured images.
Align Your Steps: This custom scheduling aligns the steps to specific key transitions, potentially providing more control over the trajectory of the diffusion process.
SD1.5: It may not show significant improvements over standard schedulers unless you're looking for precise control over the image generation.
SDXL: This can be useful for SDXL when you want more control over how different aspects of the image evolve at different stages.
Simple: A basic scheduler with a linear noise reduction curve. It's straightforward but not particularly optimized for detail or speed.
SD1.5: Suitable for very basic tasks or quick drafts. It doesn't offer much in terms of detail or complexity.
SDXL: Not recommended for SDXL, as it won't fully leverage the model's power or produce detailed images.
Normal: A more balanced version of "Simple" with slightly more nuanced transitions but still not as advanced as others like Karras or KL Optimal.
SD1.5: Decent for balanced images without complex textures. Works fine for general-purpose tasks.
SDXL: Better for simpler SDXL tasks where high detail isn't required, but otherwise, more advanced schedulers are preferred.
Beta: This refers to a beta-scheduled noise reduction process, which reduces noise based on a curve approximating a beta distribution. It offers a more gradual refinement process.
SD1.5: Useful for a slower refinement of more detailed images.
SDXL: Works well in SDXL, especially when preserving small details. The slower refinement helps retain texture and realism.
Turbo: A scheduler that sacrifices some detail for faster image generation. Noise reduction happens more rapidly in fewer steps.
SD1.5: Great for quick previews or when speed is more important than detail.
SDXL: Not recommended for SDXL, as it compromises detail and realism, two areas where SDXL excels.
Align Your Steps GITS: Similar to Align Your Steps, but emphasizing key moments during the generation process. GITS stands for "Generate In Transition Steps," offering more control.
SD1.5: This can be useful for producing more structured images but may not always lead to significantly better results than standard methods.
SDXL: More effective in SDXL for controlled and nuanced image generation, especially if you want to influence the evolution of details.
Align Your Steps 11 / 32: These are variations of the Align Your Steps method with specific settings for 11 or 32 steps. They allow for finer control over how each phase of the image generation process is handled.
SD1.5: These can be useful when you want fine control over step count, especially for generating quick drafts with decent structure.
SDXL: Works well for SDXL if you need precise control over the image evolution process, but can be slower compared to more efficient schedules like Karras or KL Optimal.
Good Combinations for SD1.5 vs. SDXL
SD1.5
For SD1.5, simpler samplers like Euler A, DDIM, or LMS can be combined with basic schedulers like:
Karras: Best for realism, even with fewer steps.
DDIM: Well-rounded for speed and detail.
KL Optimal: Produces consistent, structured results.
Turbo: If speed is your priority.
SDXL
For SDXL, where complexity and detail are key, combining advanced samplers like DDIM or DPM++ with sophisticated schedulers will yield the best results:
Karras: Excellent for photorealism and highly detailed textures.
KL Optimal: Preserves intricate details while maintaining balance.
Beta: Produces slower but detailed refinement.
Align Your Steps GITS: Good for structured transitions, especially with complex scenes.
The combination of a sampler and a scheduler, along with the number of steps, plays a significant role in the final output. But the problem with just comparing the output of these methods is that it's subjective—what's considered "good" or "bad" can vary. Instead, let's walk through the strengths of these tools and offer some useful combinations based on specific needs.
Useful Sampler/Scheduler Combos
Here's a breakdown of useful combinations for different models and scenarios.
SD 1.5 Realistic Images (High Steps > 30)
DDIM + Karras: This combo offers excellent control over detail and realism. Karras allows for smoother noise reduction across more steps, while DDIM delivers exceptional structure and fine details at a higher step count.
DPM++ 2M + KL Optimal: With KL Optimal preserving structure and DPM++ 2M offering refined noise transitions, this combination is ideal for generating sharp, photorealistic images at high step levels.
PLMS + Karras: For those seeking a balance between speed and realism, PLMS paired with Karras ensures smooth, detailed results, even with more than 30 steps.
Euler + Beta: If you need speed but still want realistic textures, Euler with Beta offers a nice blend of quick generation and gradual noise reduction.
LMS + KL Optimal: LMS excels at preserving structure, while KL Optimal ensures optimal denoising. Great for high-detail photorealism.
DPM++ 2M SDE Heun + Karras: For precision and fine detail, DPM++ 2M SDE Heun paired with Karras gives you control over texture and smooth transitions.
DPM2 A + Polyexponential: This combination is for users who want complexity and refined textures. DPM2 A works well with Polyexponential for highly detailed and dynamic images.
DPM Fast + Karras: If speed matters but you don't want to sacrifice too much quality, DPM Fast with Karras keeps the pace without losing control of the image.
DDIM + Align Your Steps 32: Align Your Steps allows DDIM to fine-tune images over a longer process, keeping realism and structure intact.
DPM++ SDE + KL Optimal: This combo leverages the SDE variant's stochastic benefits with KL Optimal's precision, offering sharp, consistent details.
SD 1.5 Realistic Images (Low Steps < 30)
Euler A + Karras: For speed and decent detail at low steps, this pairing helps you get results fast without losing too much quality.
DDIM + Beta: Beta offers slower, gradual noise reduction, which complements DDIM for cleaner images at lower steps. A good option when you need balance at lower counts.
LMS + Karras: LMS keeps the structure clean, while Karras helps maintain smoothness, making this combination work well at lower step counts.
DPM Fast + Uniform: If speed is your top priority and you're okay with losing a little precision, DPM Fast with Uniform does the job well.
DPM++ 2M + Simple: Simple yet effective, DPM++ 2M with Simple scheduling provides realistic results at low step counts.
DDIM + KL Optimal: For quick, precise results, DDIM with KL Optimal balances speed and detail, which is perfect for lower steps.
Euler + Exponential: This combination offers a faster but effective diffusion, giving you quick results with reasonable realism.
DPM++ SDE + Karras: DPM++ SDE adds texture and depth, especially when paired with Karras, for smooth, even transitions.
DPM2 A + Karras: preserves detail for more controlled low-step outputs without taking too long.
DPM++ 2S A + Turbo: When speed is the primary goal, Turbo scheduling helps generate fast results, with DPM++ 2S A adding enough structure to keep it realistic.
SDXL Realistic Images (Low Steps < 30)
DDIM + Karras: This combination balances speed and quality, allowing detailed outputs in fewer steps, perfect for SDXL's complexity.
DPM++ 3M SDE + KL Optimal: Ideal for low-step SDXL work, offering rich detail and smoothness while keeping the process efficient.
LMS + Karras: LMS ensures structural integrity, while Karras helps refine at lower step counts, making this ideal for fast SDXL outputs.
Euler A + Karras: For speed, Euler A with Karras offers a solid, fast choice for those who need lower step counts without sacrificing too much quality.
DPM++ 2M SDE Heun + Karras: This pairing ensures smoothness and depth even at lower steps, which is ideal for detailed but fast SDXL images.
DPM2 A + KL Optimal: Balances speed and complexity, perfect for low steps when you still want depth and realism.
DPM++ 2M SDE + Beta: Offers more gradual noise reduction at low steps, making it easier to retain detail while working faster.
DDIM + KL Optimal: It keeps things sharp and fast, and it is perfect for low-step SDXL outputs when precision is needed.
Euler A + Align Your Steps GITS: For those looking to control the process, Align Your Steps helps guide Euler A at low step counts.
DPM++ 2M + Polyexponential: Adds depth and balance, making it perfect for low-step photorealistic outputs in SDXL.
SDXL Realistic Images (High Steps > 30)
DDIM + Karras: This combination delivers outstanding quality for high-detail, realistic images, allowing for smooth and natural transitions in the noise reduction process. Karras ensures that even at high steps, the details are finely tuned.
DPM++ 3M SDE Heun + KL Optimal: DPM++ 3M SDE Heun works exceptionally well with KL Optimal for high-resolution, highly detailed outputs. The stochastic properties of SDE help capture intricate textures, making this pairing perfect for larger, complex images.
LMS + Karras: This pairing is excellent for creating photorealistic images with structural clarity. LMS preserves details across a large number of steps, and Karras keeps the refinement smooth and consistent.
DPM++ SDE + Beta: A versatile option for high steps, DPM++ SDE works alongside Beta scheduling to maintain smooth noise transitions while still capturing details in the final image.
DPM++ 2M SDE Heun + Karras: This pairing works especially well for more complicated images at high steps, providing detailed and structured outputs with smooth, photorealistic qualities.
DPM2 a + KL Optimal: This combination maximizes precision and texture quality, making it perfect for those longer image generations where intricate details matter most.
PLMS + Karras: A tried-and-true combination for high-resolution work, PLMS with Karras excels at creating clean, sharp images, even across many steps.
DPM++ 2M + Polyexponential: Polyexponential scheduling allows for a more complex decay curve, which, when combined with DPM++ 2M, offers better control over texture and depth in long, multi-step processes.
DDIM + KL Optimal: This pairing ensures that every step of the noise reduction process is calculated to bring out the best possible detail while working efficiently at high steps.
Euler + Beta: A solid and straightforward option, Euler with Beta is fast but still offers reasonable detail and clarity, making it a reliable choice for high-step generation when you need quicker results.
SD1.5 Illustration/Cartoon Images (High Steps > 30)
DDIM + Karras: For highly detailed and smooth cartoon or illustrative images, DDIM combined with Karras brings sharp edges and vibrant textures to life over extended steps.
DPM++ 2M + Beta: This combo allows for rich, bold colors and strong linework, making it perfect for more dynamic cartoon illustrations with high step counts.
Euler + Karras: Provides a balance between speed and fine control over stylistic elements, especially for illustrations with sharp edges and strong contrasts.
PLMS + KL Optimal: PLMS ensures consistent lines and textures, while KL Optimal guarantees detailed refinement, making it great for crisp cartoon images.
LMS + Karras: LMS is fantastic for preserving the structural integrity of lines and shapes in illustrations, while Karras adds smoother transitions between colors and details.
DPM2 a + Polyexponential: A potent combo for complex and vibrant cartoon images, especially where nuanced shading and detail are essential.
DPM++ 2M SDE Heun + Karras: This setup produces detailed cartoon work with smooth lines and vibrant colors across many steps.
Euler A + Karras: Fast and efficient, this combination works well for bold, simple illustrations at high steps, where detail needs to be maintained without sacrificing speed.
DPM++ SDE + Beta: Offers a more refined noise reduction process that works well for solid lines and softer shading in illustrations.
DPM++ 2S a + Simple: A good choice for simpler cartoon styles, offering consistent results with less complexity in the denoising process.
SD1.5 Illustration/Cartoon Images (Low Steps < 30)
Euler A + Karras: A quick and effective combination that produces bold and vibrant cartoon images with minimal steps, focusing on speed without losing too much detail.
DDIM + Simple: DDIM ensures clean lines and sharp detail, while Simple scheduling keeps the process quick and straightforward for lower step counts.
DPM Fast + Uniform: provides a speedy yet effective combo if you need fast results with vibrant colors and strong outlines.
DPM++ 2M + Karras: This pairing delivers sharp cartoon images quickly for those looking to balance speed and quality at lower step counts.
Euler A + Turbo: Prioritizes speed without losing the essential features of cartoon and illustration work. Turbo helps keep it fast while maintaining bold colors.
LMS + Simple: A good choice for low-step cartoon images when smooth lines and structure are important, but speed is still a priority.
DDIM + Karras: Works well for low-step illustration images that still require some detail and structure, with Karras smoothing the transitions.
DPM2 A + Karras: For those who want more control at lower steps, this combination ensures that the essential shapes and lines remain clean while providing smooth shading.
DPM++ 2S A + Beta: Produces clean, bold cartoon imagery quickly, beneficial for dynamic and vibrant styles.
Euler + Exponential: This combo works well for creating quick, simplified illustrations with clear lines and bright colors, optimized for low-step counts.
SDXL Illustration/Cartoon Images (High Steps > 30)
DDIM + Karras: This combination allows for sharp lines, smooth transitions, and vibrant details for highly detailed illustrations and complex cartoon work.
DPM++ 3M SDE + KL Optimal: DPM++ 3M SDE offers intricate detail and precision, while KL Optimal ensures optimal noise reduction over many steps, great for richly detailed cartoon images.
Euler + Karras: A versatile option for detailed cartoons, offering a balance between speed and control. Works well with high steps for maintaining clarity and boldness.
DPM++ 2M + Beta: It adds depth and vibrancy to illustration images while maintaining clear lines and bold shapes, which is especially effective with more steps.
DPM++ 2M SDE Heun + Karras: This combination is great for cartoon images that need detailed, refined edges and smooth color transitions, ideal for multi-step illustrations.
LMS + Karras: Preserves structural integrity in linework and shading, making it a solid choice for complex, high-step cartoon images.
DPM2 A + Polyexponential: It adds complexity and smoothness and is perfect for vibrant, detailed illustrations where you want complete control over textures and shading.
DDIM + KL Optimal: A reliable combo for detailed cartoon images with delicate shading, especially at high steps where the precision of KL Optimal comes into play.
DPM++ SDE + Karras: Smooth and reliable, this setup excels at creating well-rounded illustrations with depth and clarity at higher step counts.
Euler A + Align Your Steps GITS: Offers control over the diffusion process, ensuring clean, sharp linework and vibrant color transitions over many steps.
SDXL Illustration/Cartoon Images (Low Steps < 30)
Euler A + Karras: Quick and practical, this combination works well for bold, simple cartoon images with lower step counts, maintaining vibrancy and structure.
DDIM + Beta: Beta scheduling ensures gradual noise reduction, which is excellent for retaining color richness and clean lines at lower step counts.
DPM++ 2M + Turbo: A fast and efficient combo, perfect for quickly producing clean, dynamic cartoon images with minimal steps.
LMS + Simple: LMS ensures that linework stays sharp, while Simple scheduling keeps the process efficient for low-step illustrations.
DDIM + Karras: Effective for lower-step illustrations where detail is still important but speed is a factor, Karras smooths out transitions.
DPM Fast + Uniform: this pairing ensures that cartoon images remain bold and clear, even at lower step counts.
DPM++ 2S A + Simple: A simple yet effective combination for quick, clean cartoon images with bold lines and colors.
Euler A + Exponential: This pairing helps produce bold and vibrant images for faster results, with Exponential providing a more rapid noise decay while still retaining structure.
DPM2 A + Beta: Great for balancing speed and quality in low-step cartoon images, this combination helps maintain clean edges and vibrant colors with gradual noise reduction.
DPM++ 2M SDE + KL Optimal: Ensures smooth transitions and clarity even with lower steps, allowing for more refined, detailed cartoon images while working within a shorter timeframe.
Let me know if this was helpful for anyone and if there's anything I could do to improve the article.