Advanced settings

• Weight：

  It can control the similarity of the picture to the original cat, the default value is 1, it only exists in the model PicMate.

• CFG Scale：

  CFG scale refers to how consistent the image is with the cue——The lower value will get the more creative results; The higher value will be closer to the description text.

  Sampler：

  Sampler refers to sampling mode, i.e. diffusion algorithm de-noise sampling mode will affect its effect, different sampling mode results will be very different, the default choice “Euler a”

  Euler a: it uses the Euler method of ancestral sampling, influenced by eta parameters in sampler settings.

  Recommended Steps: 30 ~ 35

  DPM++ 2M Karras：2-order multi-step sampling method implemented in K-diffusion and using a noise schedule. Diffusers In the Hagging face are known to be the most emphatic, with the best balance of speed and quality. This "more step M" would refer more than the "single step S" in sampling, rather than the current step, thus providing better quality. but also more complex.

  Recommended steps: 35 ~ 40

  DPM++ SDE Karras: DPM++ SDE Karras: SDE version of DPM++, with noise schedule, i.e. stochastic differential equations, DPM++ was the original solver of ODE, namely the ordinary differential equations. In the version of the K-diffusion implementation, ancestral sampling is invoked in the code Method, so it is affected by the ETA parameters in the sampler settings.

  Recommended Steps: 30 ~ 35 DPM++ 2S a Karras： The 2-order single step implementation in K-diffusion, using ancestral sampling, is influenced by the eta parameter in the sampler setting. Recommended Steps: 30 ~ 35 UniPC：UniPC is a newly developed diffusion sampler in 2023 consisting of two components: a UniP and UniC. It supports any solver and noise predictor. Recommended Steps: 20 ~ 30 Steps：

  Steps refers to the iterative steps that are performed. With each iteration, AI has more opportunities to compare and adjust to the prompt and current results. It should be noted that higher iteration steps consume more computational time and cost, but do not exactly mean the better results. However, if the iteration step count is too low (less than 20), the image quality will certainly decline. seed： Random number of seeds when each image is generated. They're used as the basis for determining the initial state of diffusion. You can input seeds and keep the parameters unchanged to get the stable quality of pictures.