AI VIDEO KNOWLEDGE BASE
Diffusion Samplers and Schedulers
Compare common samplers and schedulers
01 · Samplers
Sampler comparison
| Sampler | Behavior | Work per step | Typical step range | Common use | Video models that use this as default | Notes |
|---|---|---|---|---|---|---|
| Euler | Deterministic | 1 model call | 20–35 | A simple baseline; flow-matching image and video models | Takes one direct denoising move at each step. Fast and easy to compare, but rougher than higher-order methods at very low step counts. EDM paper ↗ | |
| UniPC | Deterministic predictor-corrector | About 1 model call | 8–25 | Few-step sampling on supported models | Improves each predicted move with a correction that does not require another model call. Particularly useful when the model and schedule behave well at low steps. UniPC paper ↗ | |
| Euler a (ancestral) | Adds fresh noise | 1 model call | 20–35 | Stylized images and exploring variations | Adds noise during the run. It can create lively variation, but extra steps do not always settle toward one final image and may be less suitable for steady video. k-diffusion implementation ↗ | |
| DDIM | Usually deterministic | 1 model call | 20–50 | Reproducible images, inversion and older model workflows | A faster alternative to the original DDPM process. It remains useful, but it is not automatically the best default for newer flow-matching models. DDIM paper ↗ | |
| DPM++ 2M | Deterministic; reuses history | About 1 model call | 15–30 | Detailed still images and broadly supported workflows | — | Uses recent denoising results to make a more accurate next move without normally doubling the work. A strong image baseline when the model supports it. DPM-Solver++ paper ↗ |
| LCM | Model-specific few-step sampling | 1 model call | 2–8 | LCM, consistency-distilled, or matching accelerator models | Use only with a model or adapter trained for very few steps. Selecting LCM alone does not turn an ordinary checkpoint into a fast distilled model. LCM paper ↗ | |
| RES Multistep | Deterministic; reuses history | About 1 model call | 10–25 | Modern diffusion and flow-matching models; current video workflows | A refined exponential solver now included in ComfyUI. It is a strong modern option, but the model's published scheduler and step count still matter. RES paper ↗ | |
| DPM++ 2M SDE | Adds fresh noise; reuses history | About 1 model call | 20–35 | Natural texture and varied image results | — | The stochastic version of the multistep DPM++ approach. It can improve texture and variety, but is not a universal cure for video flicker. DPM-Solver++ code ↗ |
| DPM++ 2S a | Adds fresh noise; two-stage | About 2 model calls | 12–25 | Low-to-medium-step image generation | — | Makes two model evaluations per listed step and injects noise. It may look good at a lower step count, but it is not necessarily faster than DPM++ 2M. k-diffusion implementation ↗ |
| Heun | Deterministic correction | About 2 model calls | 15–30 | A cleaner Euler alternative when time allows | — | Checks its first move with a second model evaluation. A listed step therefore costs roughly twice an Euler step; compare model calls, not only the step number. EDM paper ↗ |
| DEIS | Deterministic; reuses history | About 1 model call | 10–30 | Fast tests on compatible diffusion models | — | An exponential-integrator method designed to reduce error at low model-call counts. It is not specifically a temporal-consistency sampler for video. DEIS paper ↗ |
| Gradient Estimation | Deterministic; reuses the last result | About 1 model call | 5–20 | Low-call experiments and supported modern workflows | — | Uses the current and previous model outputs to estimate a better direction. It is available in ComfyUI, but should be tested against the model's official default. ICML 2024 paper ↗ |
| Ralston 2S | Deterministic two-stage correction | 2 model calls | 12–25 | Quality-focused MiniMax H3 and custom-node workflows | — | Considered one of the best-looking MiniMax H3 samplers. It weights its second check more heavily and is usually supplied by a custom sampler pack, not stock ComfyUI. RES4LYF implementation ↗ |
| ER-SDE | Adds fresh noise; reuses history | About 1 model call | 15–30 | Current ComfyUI experiments where texture and variation matter | — | A newer stochastic solver that aims to keep the speed of deterministic methods while gaining the error-correcting benefit of added noise. WACV 2025 paper ↗ |
| SA-Solver | Adds controlled noise; reuses history | About 1 call; PECE costs more | 10–30 | Newer stochastic tests in current ComfyUI | — | Uses earlier results to predict and correct the next move. The PECE version performs an extra corrected evaluation, so it is slower than the standard form. SA-Solver paper ↗ |
| SEEDS-2 / SEEDS-3 | Adds fresh noise; multi-stage | 2 or 3 model calls | 8–20 | Experimental stochastic sampling at low listed steps | — | These newer ComfyUI options can use fewer listed steps, but each step evaluates the model multiple times. Judge them by total calls and output, not the small step number. SEEDS paper ↗ |
| DDPM | Adds fresh noise | 1 model call | 200–1000 historically | Research reference and workflows that explicitly require it | — | The foundational slow reverse-diffusion process. It is historically important, but rarely the practical choice for current image or video generation. DDPM paper ↗ |
Scheduler vs. Sampler
A sampler controls the calculation at each denoising step. A scheduler controls where those steps are placed across the noise range. “Euler with Simple,” for example, uses Euler for the calculation and Simple for step placement. Use the model's recommended pairing first.
02 · Schedulers
Scheduler comparison
| Scheduler | Step placement | Common use | Video models that use this as default | Notes |
|---|---|---|---|---|
| Normal | Evenly spaced model timesteps | General use when the model or workflow specifies normal | — | Follows the model's native timestep range at even intervals. It is a neutral baseline, not a universal default. ComfyUI implementation ↗ |
| Simple | Even samples from the model's stored noise levels | Many current flow-matching and video workflows | Selects evenly from the noise schedule already stored with the model. Use it when the published workflow calls for simple. ComfyUI implementation ↗ | |
| SGM Uniform | Even model timesteps using the SGM convention | SGM-style and model-specific workflows | — | A slightly different form of uniform timestep spacing. It should not be assumed equivalent to normal for every model. ComfyUI implementation ↗ |
| Karras | More steps near the low-noise end | Compatible image models where final detail needs more attention | — | Spends more of the step budget near the end of denoising. Common with sigma-based image workflows, but not ideal for every modern model. EDM paper ↗ |
| Exponential | Even spacing on a logarithmic noise scale | Sigma-based model and sampler experiments | — | Reduces the noise level by a consistent ratio rather than a consistent amount at each step. k-diffusion implementation ↗ |
| DDIM Uniform | Even samples from DDIM training timesteps | DDIM and older workflows designed around its timestep layout | Keeps the timestep selection expected by DDIM-style sampling. The name refers to the schedule, not the DDIM sampler itself. ComfyUI implementation ↗ | |
| Beta | More steps near the beginning and end | Low-step tests on compatible image and video models | Uses a beta distribution to devote less of the step budget to the middle of the noise range. Results remain model-dependent. Beta Sampling paper ↗ | |
| Linear Quadratic | A linear section followed by a curved section | Mochi and other workflows that explicitly recommend it | Originally adopted from a Mochi video workflow. Treat it as a model-specific option rather than a general upgrade. ComfyUI implementation ↗ | |
| Manual Sigmas | Exact noise levels supplied by the workflow | Models with a hand-tuned few-step schedule | This is an explicit list of noise levels rather than a scheduler preset. LTX-2.5 uses different manual lists for its main and upscale passes. LTX-2.5 workflow ↗ | |
| KL Optimal | Curved spacing across the noise range | Specialized workflows and controlled comparisons | — | Uses a mathematically derived curve to place noise levels. Prefer the model author's recommendation unless you are testing schedules directly. ComfyUI implementation ↗ |