A Stable Diffusion sampler is the method that removes noise at each step (Euler, Euler a, DPM++ 2M, DDIM), while the scheduler, shown as ‘Schedule type’ in AUTOMATIC1111, sets how much noise is left after each step (Karras, Exponential, SGM Uniform); ‘DPM++ 2M Karras’ is simply one of each. For SDXL, SD 1.5, and fine-tuned models, start with the sampler and scheduler the model card recommends, then change only one at a time.
Sampler vs scheduler: the short answer
Stable Diffusion makes an image by starting from noise and removing it over a number of steps. Two settings control that loop. The sampler is the method used to compute each denoising step: Euler, Euler a, Heun, DPM++ 2M, DDIM, UniPC, and so on. The scheduler, which AUTOMATIC1111 labels ‘Schedule type’, sets the noise level the image should reach after each step, so it decides how the steps are spaced between very noisy and clean.
ComfyUI's KSampler node shows the split directly: its documentation describes sampler_name as the sampling algorithm, which affects speed, style, and quality, and scheduler as the setting that controls how noise is gradually removed. A combined name such as ‘DPM++ 2M Karras’ is one sampler (DPM++ 2M) plus one schedule (Karras).
Hugging Face Diffusers uses the words differently. There, a ‘scheduler’ is one object that holds both the update rule and the noise schedule: DPM++ 2M is DPMSolverMultistepScheduler, and Karras is an option on it (use_karras_sigmas=True). Same ideas, different labels.
What ‘Schedule type’ means in AUTOMATIC1111
AUTOMATIC1111 version 1.9.0 moved scheduler selection into the main UI. Since then the txt2img and img2img tabs show two dropdowns, ‘Sampling method’ and ‘Schedule type’, next to ‘Sampling steps’ (default 20). Older names that combined both, such as ‘DPM++ 2M Karras’ in saved generation settings, are split into the sampler ‘DPM++ 2M’ plus the schedule type ‘Karras’ when they are loaded.
The default schedule type is ‘Automatic’, which uses the sampler's own preferred schedule. In the web UI's sampler table, DPM++ 2M defaults to Karras and DPM++ 2M SDE to Exponential, so ‘DPM++ 2M’ with ‘Automatic’ is the old ‘DPM++ 2M Karras’. The other schedule types are Uniform, Karras, Exponential, Polyexponential, SGM Uniform, KL Optimal, Align Your Steps, Simple, Normal, DDIM, and Beta.
ComfyUI's KSampler offers a similar list under its own names: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, normal, linear_quadratic, and kl_optimal, alongside samplers such as euler, euler_ancestral, heun, dpmpp_2m, dpmpp_2m_sde, dpmpp_3m_sde, ddim, uni_pc, and lcm.
Sampler names decoded: Euler, ‘a’, 2M, SDE, and Heun
Euler takes one simple step per model call; k-diffusion implements it as the Euler steps described by Karras et al. (2022). Heun, from the same paper, corrects each step with a second model call, so it uses about twice as many model calls per step. DPM++ 2M is the multistep DPM-Solver++ method: one model call per step, reusing the previous step's result to improve accuracy. The DPM-Solver++ paper reports high-quality guided samples in 15 to 20 steps.
An ‘a’ means ancestral. Ancestral samplers such as Euler a, DPM2 a, and DPM++ 2S a add fresh random noise after each step; the k-diffusion code calculates both how far to step down and how much new noise to add. ‘SDE’ variants such as DPM++ SDE and DPM++ 2M SDE are stochastic as well. Plain Euler, Heun, and DPM++ 2M add no new noise after the starting latent with their default settings.
That difference shows up when you change the step count. With a deterministic sampler, the only randomness is the starting noise from the seed. With an ancestral or SDE sampler, the result also depends on the noise injected at every step, so changing the number of steps changes that noise and can change details or even the composition.
Scheduler names decoded: Karras, Exponential, SGM Uniform, and Align Your Steps
Karras is the noise schedule from Karras et al. (2022). The paper spaces the noise levels so the steps get shorter as the noise gets lower (rho = 7). More of the step budget goes to the final low-noise steps, where fine detail is resolved, and the early high-noise steps are larger. Karras is applied at sampling time, and AUTOMATIC1111 uses it by default for DPM++ 2M with ordinary SD 1.5 and SDXL checkpoints (the Diffusers guide is more cautious and suggests it only for models trained with Karras sigmas). The exception is few-step distilled models, which need the schedule their card names.
Diffusers maps the other common names to its own options: Exponential to exponential sigmas, Beta to beta sigmas, and both SGM Uniform and Simple to ‘trailing’ timestep spacing. SGM Uniform gets its own section below.
Align Your Steps is an optimized schedule from NVIDIA research; Diffusers ships a 10-step version for SDXL and says it can produce a high-quality image in 10 steps. A schedule never changes the model itself. It only changes where the steps fall between full noise and a clean image.
PhoneDiffusion generates on-device on iPhone or iPad, with optional cloud models. Download on the App Store →
What is SGM Uniform? SGM Uniform vs Simple, Normal, and Karras
SGM Uniform (sgm_uniform in ComfyUI) is a noise schedule that spaces the steps evenly in timesteps, the model's training clock that runs from 999 (almost pure noise) down to 0, instead of spacing them by noise level. In ComfyUI's code it is the ‘normal’ schedule with one change: it splits the range from the highest to the lowest timestep into steps + 1 even points, drops the last point, and then ends at zero noise. AUTOMATIC1111's ‘SGM Uniform’ schedule type uses the same calculation. Diffusers has no option with that name and maps it to timestep_spacing="trailing", which its guide illustrates with ten timesteps: 999, 899, 799, and so on down to 99.
The name matches sgm, the Python package of Stability AI's generative-models repository, which the SDXL base model card links as its reference code. That repository's legacy DDPM discretization also picks roughly evenly spaced timesteps that count down from the top and leave out the last one.
SGM Uniform vs Simple: in practice they are almost the same. Simple steps through the model's own table of training noise levels at even intervals, so on a model with 1,000 training timesteps, such as SDXL base, the two land on nearly identical noise levels, and Diffusers maps both to trailing spacing. Normal also spaces timesteps evenly, but it keeps the lowest training timestep as its final stop, so its last stop before zero is the model's lowest noise level (about 0.03 on SDXL) instead of about 0.34.
Karras is the real contrast. As a worked example, calculated from the ComfyUI code with the noise schedule in SDXL base's scheduler config (beta 0.00085 to 0.012 over 1,000 timesteps) at 10 steps: SGM Uniform and Simple both start at a noise level (sigma) of about 14.6 and stop at about 0.34 before the final jump to zero, while Karras (rho = 7) runs down to about 0.03 and puts five of its ten stops below 1.0, against three for SGM Uniform. Karras spends more of the budget on fine detail at the low-noise end; SGM Uniform keeps the steps evenly spread over the model's training timesteps.
When to use SGM Uniform: when the model card asks for it. SDXL-Lightning's card says to use the Euler sampler with the sgm_uniform scheduler in ComfyUI, and Euler with trailing timesteps in Diffusers; Hyper-SD's Diffusers examples for its 2-, 4- and 8-step SD 1.5 and SDXL LoRAs set DDIM with trailing spacing. Diffusers adds that trailing spacing typically gives more detail at low step counts, while the difference is less obvious at standard step counts. On an ordinary SD 1.5 or SDXL checkpoint at the default 20 steps, treat SGM Uniform as one schedule to test against Karras with the same seed, not an upgrade.
Best sampler and scheduler for SDXL, SD 1.5, and fine-tuned models
There is no single winner, and the maintainers say so: Diffusers advises starting from the checkpoint's default and swapping only when you need a different speed or quality tradeoff. For many latent diffusion checkpoints it names DPM++ 2M SDE Karras as a strong all-purpose option, and it notes that Euler or Euler a often work well for anime-style images.
Check the model card first. The SDXL base 1.0 repository ships an Euler scheduler (EulerDiscreteScheduler) with leading spacing and Karras sigmas turned off. Illustrious XL, an illustration-focused SDXL model, recommends Euler a at 20 to 28 steps with CFG 5 to 7.5. The same rule applies to SD 1.5 fine-tunes: the card for the exact checkpoint beats any general ranking.
Few-step distilled models are where the card matters most. SDXL-Lightning's card requires Euler with ‘trailing’ timesteps in Diffusers and the Euler sampler with the sgm_uniform scheduler in ComfyUI, and Diffusers pairs its LCM scheduler with LCM models or LoRAs. Follow those instructions before experimenting.
How to compare samplers and schedulers fairly
Fix the prompt, negative prompt, seed, size, CFG, and model. Start at 20 steps, the default in both AUTOMATIC1111 and ComfyUI's KSampler (how many steps you need depends on the model). Then change only one thing: first the sampler with the schedule left on its default, then the schedule with the sampler fixed.
Repeat the two best combinations at a higher step count, such as 30, before deciding. Expect ancestral and SDE samplers to move more between those runs, because they receive new noise at every step. Keep the comparison to two or three candidates so each difference stays easy to see.
Record the winning sampler and schedule with the seed and the rest of the txt2img settings. The same names in a different app are not guaranteed to produce the same pixels, because implementations, default schedules, and numerical precision differ.
PhoneDiffusion availability and article scope
PhoneDiffusion provides text-to-image and separate editing tools on supported iPhone and iPad devices running iOS 17 or later, using curated Core ML model packs installed through the app. In the current release, each model pack carries its own sampler and step-spacing setting as part of its defaults, chosen for that pack, instead of a desktop-style sampler dropdown. For example, the 8-step Hyper packs, such as Stable Diffusion XL Base Hyper and Animagine XL Hyper, run DDIM with trailing spacing, the Diffusers option SGM Uniform maps to, as the Hyper-SD card instructs; the Stable Diffusion 1.5 and 2.1 base packs and the RealVisXL V5 Lightning and DreamShaper XL Lightning packs use DPM++ SDE with Karras spacing. You cannot switch a pack to another schedule. Optional cloud models (paid monthly or yearly plans) generate online and expose no sampler or step settings.
PhoneDiffusion does not provide a general production importer for arbitrary checkpoints or LoRAs, and the AUTOMATIC1111 and ComfyUI names above describe desktop tools. A desktop recipe, sampler name, or seed is not promised to reproduce identical pixels on iPhone.
Field notes
- Sampler = how each denoising step is computed; scheduler (‘Schedule type’) = how much noise is left after each step.
- ‘DPM++ 2M Karras’ is the DPM++ 2M sampler plus the Karras schedule; AUTOMATIC1111 1.9.0 and later show them as two dropdowns.
- Ancestral (‘a’) and SDE samplers add new noise every step, so results shift more when the step count changes.
- SGM Uniform spaces steps evenly in timesteps and leaves out the lowest timestep before ending at zero noise; it is nearly identical to Simple, maps to ‘trailing’ in Diffusers, and is the schedule SDXL-Lightning asks for in ComfyUI.
- Start with the model card's recommendation or the checkpoint default; few-step models need the exact sampler and schedule they were built for.
Try it on your phone
- Read the model card and use its recommended sampler, schedule, steps, and CFG.
- Fix the seed, prompt, size, and CFG, and generate at 20 steps.
- Change only the sampler, then only the schedule, and compare the results side by side.
- Repeat the best two combinations at a higher step count and save the winner with its seed.
Sources and verification
Primary and maintainer documentation checked for this guide. Product-specific details were reviewed against the current PhoneDiffusion scope.
- Hugging Face Diffusers: Schedulers API overviewMaintainer reference defining a Diffusers scheduler and mapping A1111/k-diffusion sampler names (DPM++ 2M, Euler a, …) and schedule types (Karras, sgm_uniform, simple, exponential, beta) to Diffusers schedulers and options.
- Hugging Face Diffusers: Schedulers guideMaintainer guide: start from the checkpoint default, DPM++ 2M SDE Karras as an all-purpose option, Euler/Euler a for anime, LCMScheduler for LCM models, timestep spacing (leading, linspace, trailing), Karras sigmas only for models trained with them, and the 10-step Align Your Steps schedule.
- k-diffusion: sampler implementations (k_diffusion/sampling.py)Katherine Crowson's reference code for Euler, Euler ancestral, Heun, DPM2, DPM++ 2S a, DPM++ SDE, DPM++ 2M and the Karras/exponential noise schedules; shows ancestral steps adding fresh noise and Heun's second model call per step.
- Karras et al. (2022): Elucidating the Design Space of Diffusion-Based Generative ModelsThe NVIDIA paper behind the ‘Karras’ noise schedule and the Euler/Heun sampling algorithms used in k-diffusion.
- Lu et al. (2022): DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic ModelsThe paper behind the DPM++ samplers, reporting high-quality guided samples in 15 to 20 steps.
- AUTOMATIC1111 Stable Diffusion web UI: changelog (1.9.0: scheduler selection in main UI)Project changelog recording scheduler selection in the main UI in 1.9.0 and the later Align Your Steps, KL Optimal, Normal, DDIM, Simple, and Beta schedulers.
- AUTOMATIC1111 web UI: Sampling method and Schedule type controls (processing_scripts/sampler.py)Source of the ‘Sampling method’, ‘Schedule type’, and ‘Sampling steps’ (default 20) controls on the txt2img and img2img tabs.
- AUTOMATIC1111 web UI: k-diffusion sampler list and default schedules (sd_samplers_kdiffusion.py)Sampler table showing each sampler's default schedule (DPM++ 2M: Karras, DPM++ 2M SDE: Exponential), used when Schedule type is ‘Automatic’; the schedule list is in modules/sd_schedulers.py.
- AUTOMATIC1111 web UI: schedule types (modules/sd_schedulers.py)Source of the ‘SGM Uniform’, ‘Simple’, ‘Normal’ and ‘Karras’ schedule types: sgm_uniform spaces steps + 1 timesteps evenly from the highest to the lowest noise level, drops the last one and appends zero noise.
- ComfyUI: KSampler nodeOfficial node reference: seed, steps (default 20), cfg, sampler_name (the sampling algorithm), scheduler (how noise is gradually removed), and denoise.
- ComfyUI repository: sampler and scheduler names (comfy/samplers.py)Source of the KSampler sampler names (euler, euler_ancestral, dpmpp_2m, …) and scheduler names (simple, sgm_uniform, karras, exponential, ddim_uniform, beta, normal, linear_quadratic, kl_optimal). SCHEDULER_HANDLERS defines sgm_uniform as normal_scheduler with sgm=True (evenly spaced timesteps, last point dropped) and simple as even steps through the model's own sigma table.
- Stability AI generative-models (package sgm): discretizer.pyReference code linked from the SDXL base model card; its legacy DDPM discretization picks roughly equally spaced timesteps counting down from the top, without the final one.
- Stability AI: SDXL base 1.0 scheduler configurationThe checkpoint's default scheduler: EulerDiscreteScheduler, leading timestep spacing, use_karras_sigmas false.
- ByteDance: SDXL-Lightning model cardModel card requiring Euler with ‘trailing’ timesteps in Diffusers and the Euler sampler with the sgm_uniform scheduler in ComfyUI.
- ByteDance: Hyper-SD model cardModel card whose Diffusers examples set DDIMScheduler with timestep_spacing="trailing" for the 2-, 4- and 8-step SD 1.5 and SDXL LoRAs.
- OnomaAI Research: Illustrious XL model cardModel card for an illustration-focused SDXL model recommending Euler a, 20 to 28 steps, and CFG 5 to 7.5.