Stable Diffusion steps are the number of denoising passes that turn random noise into the image. For a regular SD 1.5 or SDXL checkpoint, 20 to 30 steps is the usual range: 20 is the default in AUTOMATIC1111 and ComfyUI, and the DPM-Solver++ authors report high-quality guided samples in 15 to 20. Fast LCM and Lightning models are built for 1 to 8 steps. More steps cost time and stop helping after a point.
How many steps do you need: the short answer
Use the step count the model was made for. For a regular Stable Diffusion 1.5 or SDXL checkpoint with a modern sampler such as DPM++ 2M, start at 20 steps and try 30 if fine detail still looks unfinished. Fast distilled models are different: an LCM model needs about 2 to 8 steps, and SDXL-Lightning checkpoints are trained for exactly 1, 2, 4, or 8.
Each step is another pass through the model, so the time per image grows with the step count. Diffusers says more denoising steps usually lead to higher quality ‘at the expense of slower inference’, and the gain shrinks once the image has settled. On a phone, that trade matters more: twice the steps means roughly twice the wait.
What a step actually does
Stable Diffusion starts from random noise in a compressed latent image and removes it a little at a time. One step is one pass of the denoising model (two for samplers such as Heun) plus one update by the sampler. The scheduler decides how much noise should be left after each step, so 20 steps means the path from pure noise to a clean image is divided into 20 stops.
Fewer stops make each jump larger. Too few for the sampler and the image can stay noisy or unfinished; enough and the remaining changes become small refinements. The sampler matters here: the samplers and schedulers guide explains why some samplers finish in fewer steps than others.
Steps for SD 1.5, SDXL, and fine-tuned checkpoints
AUTOMATIC1111's Sampling steps and ComfyUI's KSampler both default to 20, while the Diffusers pipeline defaults to 50. The DPM-Solver++ paper, the method behind the DPM++ samplers, reports high-quality guided samples in 15 to 20 steps, so 20 to 30 is a sensible range to test with those samplers.
Fine-tuned checkpoints often state their own range, and the model card beats any general rule. Illustrious XL, for example, recommends Euler a at 20 to 28 steps with CFG 5 to 7.5. If a card gives a range, generate at both ends with the same seed and keep the lower count if you cannot see a difference.
PhoneDiffusion generates on-device on iPhone or iPad, with optional cloud models. Download on the App Store →
Fast models: LCM and Lightning
Distilled models are trained to finish in a handful of steps. The LCM-LoRA card says it reduces the number of inference steps to between 2 and 8, with guidance off or between 1.0 and 2.0. SDXL-Lightning ships separate 1-, 2-, 4-, and 8-step checkpoints, calls the 1-step model experimental, and says to use the same number of steps as the loaded model with guidance turned off (guidance_scale=0 in Diffusers; CFG 1 in ComfyUI or AUTOMATIC1111).
Treat those numbers as the recipe, not a minimum. Raising a Lightning or LCM model to 30 steps, or putting CFG back to 7, does not turn it into a slow, high-quality model; follow the card's steps, sampler, and guidance setting first, using your tool's equivalent value for guidance.
Steps in img2img and inpainting
In image-to-image and inpainting, strength (denoising strength) decides how many of the steps actually run. Diffusers gives the example of 50 steps at strength 0.8: noise is added for 40 steps and the image is denoised for 40. A low strength runs few steps and keeps most of the original; a high strength runs nearly all of them.
AUTOMATIC1111 works the same way by default. Its option ‘With img2img, do exactly the amount of steps the slider specifies’ is off, because img2img normally runs fewer steps at lower denoising. So if a low-strength edit looks unfinished, raising the total steps gives the edit more steps to work with.
How to find the right step count for your model
Fix the prompt, seed, size, sampler, and CFG, then generate the same image at a few step counts, for example 12, 20, and 30 for a regular checkpoint. Compare them at full size. Pick the lowest count where faces, edges, and textures stop improving; anything above it only costs time.
If an image has a specific problem, more steps rarely fix it. Wrong anatomy, distorted faces, misspelled text, or a prompt detail the model ignored usually need another seed, a clearer prompt, or a targeted edit instead. Keep the seed fixed while you test; see Stable Diffusion seeds explained.
PhoneDiffusion availability and article scope
PhoneDiffusion runs curated Core ML model packs (SD 1.5, SD 2.1, and SDXL) on supported iPhone and iPad devices with iOS 17 or later; local generation is private and works offline once a pack is downloaded. It also offers optional cloud models, such as FLUX and Seedream, on paid monthly or yearly plans; those are text-to-image only, and the text prompt is sent online to generate the image.
Each local pack comes with its own default step count, sampler, and CFG. In the current release, Stable Diffusion 1.5 and 2.1 start at 20 steps, while the fast packs start lower: Stable Diffusion XL Base Hyper and Animagine XL Hyper at 8, DreamShaper 8 LCM at 8, RealVisXL V5 Lightning at 5, and DreamShaper XL Lightning at 4. With PhoneDiffusion Pro you can change Steps in Parameters from 1 to 30; Reset returns to the pack's defaults.
For image-to-image and Inpaint, the Transformation or Replacement strength shows how many steps will run, for example ‘12 of 20’. If you give a result a thumbs-down, some of the app's tips note that more steps are not a guaranteed improvement and that adding steps alone does not reliably fix anatomy. Desktop settings such as AUTOMATIC1111's img2img step option and the Diffusers defaults above describe other tools.
Field notes
- A step is one denoising pass; more steps take proportionally longer.
- Regular SD 1.5 and SDXL checkpoints usually need about 20 to 30 steps with DPM++ samplers; start at 20.
- LCM and Lightning models are built for 1 to 8 steps; use the count on their model card.
- In img2img and inpainting, strength decides how many of the steps run.
Try it on your phone
- Read the model card and start with its recommended steps, sampler, and CFG.
- Fix the prompt, seed, size, and CFG.
- Generate the same image at a few step counts, such as 12, 20, and 30.
- Compare at full size and keep the lowest count where detail stops improving.
- Fix specific errors with a new seed, a clearer prompt, or an edit instead of more steps.
Sources and verification
Primary and maintainer documentation checked for this guide. Product-specific details were reviewed against the current PhoneDiffusion scope.
- Hugging Face Diffusers: StableDiffusionPipeline num_inference_stepsAPI reference: num_inference_steps defaults to 50; more denoising steps usually lead to a higher-quality image at the expense of slower inference.
- 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 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.
- ComfyUI: KSampler nodeOfficial node reference: seed, steps (default 20), cfg, sampler_name (the sampling algorithm), scheduler (how noise is gradually removed), and denoise.
- 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.
- 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.
- LCM-LoRA for Stable Diffusion v1.5: model cardModel card: LCM-LoRA reduces the number of inference steps to between 2 and 8; disable guidance_scale or keep it between 1.0 and 2.0.
- SDXL-Lightning: model card: step countsModel card: 1-, 2-, 4-, and 8-step distilled checkpoints; use the same number of inference steps as the loaded model with guidance off (guidance_scale=0 in its Diffusers code); the 1-step model is experimental.
- Hugging Face Diffusers: Image-to-image: strength and stepsMaintainer guide: strength sets how many of the steps run; 50 steps at strength 0.8 adds noise for 40 steps and denoises for 40.
- AUTOMATIC1111 Stable Diffusion web UI: shared_options.py (img2img steps)Source: ‘With img2img, do exactly the amount of steps the slider specifies’ is off by default, because img2img normally runs fewer steps at lower denoising.