SoloMix版本v1.0 (ID: 358688)

This is a checkpoint that is very friendly to TI for generating realistic style characters.

Most TIs that are correctly trained and produced based on the original v1.5 base model can achieve good or even better restoration on this model.

Many LoRAs can also achieve good results on this checkpoint.(Based on my limited testing experience, some LoRAs trained on chilloutmix or its derivatives may not work well on this model.)

I am not a professional model maker, and I am not even a good image generator.

The reason for merging such a v1.5-based mix model at this time comes from my long-standing attempts at TI training.

In the process of training TI, I found that the final checkpoint used also has a great impact on the performance of TI, in addition to the training itself.

I encountered great obstacles when trying to train TI based on East Asian faces. I tried hard for a long time, but I was never able to reproduce my previous success in training TI for Caucasian faces.

Back then, it was very difficult to find TIs trained on East Asian celebrity faces on Civitia that could accurately restore the characters. Even now, it is still a rare sight.

This checkpoint performs well on both Caucasian and East Asian faces.

Since I haven't been able to successfully train character TIs on SDXL models, I've been stuck on version 1.5. I hope the situation of TI can be improved in future versions of Cascade or stabede diffusion v3.

If the performance of this checkpoint brings you satisfaction or surprise, then it is all thanks to the various checkpoints mixed in here. For example:ICBINP, epiCPhotoGasm, DreamShaper, RealCartoon-Realistic, The Truality Engine, Realistic Vision, Beautiful Realistic Asians... Thanks to the creators of these models, the credit goes to them.

In terms of usage, there is nothing special to note.

Creating engaging content takes effort, which I'm not good at. I usually rely on ChatGPT for this.

Negative prompts can also achieve good results with TI alone, such as:easynegative, FastNegativeV2, JuggernautNegative-neg, epiCPhoto-neg... and so on. Use them alone or mix 'em up is OK.

I'm used to using the long list of prompts in my example.

That's all.

Testing out this checkpoint brought back the joy when I first discovered stable diffusion.

Wish U have fun!

描述:

训练词语: