Euler a vs DPM++ 2M Karras: when the sampler matters

Samplers are different numerical methods for walking from noise to an image. They mostly matter in two ways: whether the result converges, and how many steps you need. Ancestral samplers (Euler a, DPM++ 2S a, and anything with "a") They add a little fresh noise at every step. Results are lively and varied, but they never settle: change the step count and the image changes. Good for exploration, bad for "I liked seed 1234 at 20 steps, let me do 30 for detail". Convergent samplers (Euler, DPM++ 2M, DPM++ 2M Karras) They settle on one image as steps increase. DPM++ 2M Karras is a reliable default at 20-30 steps. "Karras" refers to the noise schedule from Karras et al. (2022), which spends more steps at low noise where fine detail is decided. LCM sampler Only for LCM models or the LCM-LoRA. Do not use it with an ordinary checkpoint, and do not use other samplers with LCM unless the model card says so. Method 1. Fix prompt, seed, size, CFG. 2. Render Euler a at 20 and 30 steps: note that they differ. 3. Render DPM++ 2M Karras at 20 and 30: note that they are nearly the same. That five-minute test explains half the "I can't reproduce my own image" threads. Reproducing someone else's image You need model, sampler, scheduler, steps, CFG, seed, size, and any LoRAs with weights. Missing any one, and an ancestral sampler especially, you will get a cousin of the image, not the image.