Best workflow so far on SD 3.5, the newly released Stable Diffusion 3.5 best workflow by Alien Machine AI is here! I tried keeping things to a minimal but with some features so you can get started fast but it's not too boring either.

You will end up with two slightly varied images by running my "SD 3.5 Best Workflow - Large fp8 scaled Workflow - Two Stage Sampling" type ComfyUI Workflow.

You will just need to access the KSampler (Efficient) Node in your Comfy install which has several features. It allows the conditioning positive/negative, VAE, model, etc. to pass through and easily connect in a daisy chain. The VAE is handled inside the individual node, you don't need the separate VAE decode. It should also show a preview of the image as it is being denoised.

You will need the triple clip loader to load clip_l.safetensors, clip_g.safetensors and t5xxl_fp16.safetensors or the fp8 (depending on RAM, not VRAM) files.

I am using the SD3.5_large_fp8_scaled.safetensors file, and Stable Diffusion 3.5 is working on NVIDIA 3090!

VAE is baked.

Inspired from official workflow specification from ComfyUI regarding Stable Diffusion 3.5 and from trial and error.