Hello everyone,
Many of you may already know me, so I'll skip the introduction and get straight to the point.
Extreme Realism SD3
For those who use SD3 medium, you're likely aware of the numerous issues it has. I've attempted to fix these problems, but to be honest, it's quite challenging. Additionally, SD3 medium lacks realism. Therefore, I decided to train a LoRA model, which has shown promising results. Please note that this model is still experimental, so occasional issues may arise, though the probability is low.
Recommended Parameters:
LoRA Weight : 0🆙1
VAE : No Need
Sampler : DPM++ 2M SGM Uniform
Steps : 20➡30
CFG : 3➡4
Upscaler : R-ESRGAN 4x+
If you encounter any issues, I recommend using ComfyUI for a better experience. Here's the workflow: ComfyUI Workflow. Open the link, select the LoRA model, choose the LoRA strength, and hit the run button.
Upcoming Update:
Recently, Stability AI updated the SD3 license, giving us more freedom to work with it. I have already started working on SD3, and everything is almost set up. I plan to announce and hopefully publish our fine-tuned SD3 model later this month or in early August.
