Lz-SDXL model has undergone fine-tuning training for about half a month in four versions, and currently the latest version 4.0 is uploaded
The version looks quite good, but there are still many areas that need improvement. XL has high requirements for configuration, and the training time for fine-tuning is quite long. Fortunately, the basic model is excellent, and the final product is still quite good. LORA has not been tested yet, and we hope everyone can provide more feedback.
The prompt word XL also has a good effect on natural language on TAG, and everyone can try it more.
If you are using it for the first time, you can use the recommended prompts below.
TAG example: general photography of a shooting beautiful cyborg female, rainbow hat, Huge break (Beauty and detailed arm) Embraced, delete white fly, enrich fly, glowing, level highly detailed, enrich detail, face symmetry, masterpiece, await winning Sharp focus, concept art. low key,&k, uhd, low key, Octane rendec
Negative: (worst quality low quality, cgi bad eye, worst eye, illustration, carton). modified, destroyed, displayed., collectively drawn. bad
Anatomy, wrong Anatomy
Recommended dimensions: 1024x768, 1024x1536
It is recommended to speak up with great resolution, and it is not recommended to enable high-resolution repair
Sampler: DPM++3M SDE Karras
CFG scale: 3-7 (It is recommended to use BCFG when creating creative compositions, which may have unexpected effects. The higher the configuration ratio, the more realistic it may be, depending on the prompts, samplers, and steps)
VAE: built-in
No need to activate facial repair