My personal Stable Diffusion 3 workflow. Everything is tested and all settings adjusted to produce the best results.

Download the model and text encoders:  https://huggingface.co/stabilityai/stable-diffusion-3-medium

6/16/2024

• Some LoRA were released, so I added a LoRA stack.
• I added this contrast tweaked VAE to stage one, I prefer how it looks.  https://civitai.com/models/512311/simple-sd3-vae-color-tweaks
• Some settings tweaks were made to the 2.0x upscale stage, so that you can't see the seam.
• Added ModelSamplingSD3 back in. It wasn't working in my previous tests, but an update may have fixed it.

6/14/2024

• I noticed that SD3 has issues when going over 1536 resolution, this seems to be the max resolution for SD3. I suspect this could change in the future, since SD 1.5 and SDXL don't have this limitation. It may require a custom SD3 text encoder node, that accepts dimensions. Because of this, I switched to using the Ultimate SD Upscale for 1.5x. Perp-Neg was better for upscaling, because it's better than standard and can use subject masking, but it only allowed up to 1536x1536 resolution images before it introduced artifacts.
• Removed subject masking, because Ultimate SD Upscale accepts an image and not a latent image with a mask applied.
• Changed to 5.0 cfg scale, this seems to be the sweet spot.

6/13/2024

• Added background and subject clipseg masking to upscale to 1.5x using Perp-Neg Guidance.
• Added face detailer.

6/12/2024

• Initial release. PAG and Kohya Deep Shrink do not work for SD3 (yet). Perp-Neg Guidance does work well at 50%.