Sal Vibrant Realism版本v4.0 (ID: 199116)

The model is meant to give out realistic output as well generate less dull outfit and poses. It can work with fantasy generation as well. Targeting for a light weight fast generation with low steps and CFG scale requirement.

Non-Technical Usage Guide:
Sampling method: DPM++ 3M SDE Karras, DPM++ 2M Karras, Restart

width and height: anything in range of 512 ~ 1024

Sampling steps: 20 minimum

CFG scale: 2 ~ 4 ( Some generation can go up to 5, any higher will start to look like 3D render )

hires fix:
upscaler: 4x-UltraSharp
upscale: 2 minimum
denoise strength: 0.35 ~ 0.5

Clip skip: 1 ~ 2 ( 3 and above it started to not picking up what we prompted )
(It was trained on clip skip 1 but for some reason clip skip 2 often yield better output)

Suggested VAE if necessary: vae-ft-mse-840000-ema-pruned

This model is very sensitive to LoRA so please keep that in mind. It will require to boost up the steps and CFG scale to get a good outcome with LoRA.

Avoid using these negative prompt as it will messed up the proportions:
(worst quality, bad quality, low quality, normal quality)

These are some prompt that I have discovered that will messed up the generations. I'll update it accordingly.

描述:

Added difconsistencyRAWVAE_v10MEDIUMSafestensor
Added more dataset

训练词语: