Sampling Methods
Sampling methods determine how the diffusion model converts random noise into coherent images. Different samplers offer various trade-offs between speed, detail, creativity, and coherence.
- Euler: Fast sampler with a distinctive look, good for artistic styles.
- Euler a (Ancestral): Adds controlled randomness for more creative, varied results.
- Heun: High-quality sampler that produces detailed results but runs slower.
- DPM++ 2M: Balanced sampler with good detail and reasonable speed.
- DPM++ SDE: Adds stochastic elements for more variation in outputs.
- DDIM: Fast and deterministic with consistent results for the same seed.
- PLMS: Efficient sampler that works well with fewer steps.
- LMS: Simplified sampler that can produce good results with the right settings.
Samplers should be chosen based on your specific goals. For exploration and discovering unexpected creative possibilities, ancestors samplers like Euler a or DPM++ SDE introduce beneficial randomness. For precise, controlled results or when matching existing images, deterministic samplers like DDIM or DPM++ 2M provide more consistent outputs. Many artists develop preferences for particular samplers that complement their aesthetic style or workflow.