Warning: Version 1.0 and all other variations of version 1.0 are experimental.
If you notice low quality, please report it.
About Version 1.0-Crystal
Version 1.0-Crystal is the most complete and best variation of version 1.0.
It features the highest quality and excellent prompt following capability.
This version is available in two formats: fp16 and fp32.
For higher quality and better details, I highly recommend using the fp32 version.
Unfortunately, version 1.0-Orange suffered from data bias issues.
Good support of LoRA.
Version 1.0-Crystal includes numerous significant improvements, particularly in the UNet and text encoder components.
Recommended settings for version 1.0-Crystal:
Sampling Method: LCM
Steps: More than 8 steps (recommended: 10 steps)
CFG Scale: 1.0–1.5 (1.5 is preferred)
CLIP Skip: 1
Hires Fix: If you have access to Hires Fix, please use it. Otherwise, make sure to refine the image generated by txt2img using img2img. For refining (img2img), use a Denoising Strength between 0.4 and 0.6.
Please report any low-quality images you encounter during image generation.
Thank you, YSNRFD
About Version 1.0-Orange
This Version have better VAE (baked)
Good support of LoRA (enhanced)
CLIP model enhanced
better UNet pruning
Settings:
Sampling method: LCM
CFG scale: 1.0-1.5
Steps: 6<X
CLIP skip: 1.0
Mid-Range Anime Style Checkpoint