I’ve been using other people’s presets forever, and the nodes were all tangled up like a pile of noodles—I didn’t dare touch a thing. After reading this tutorial on building SDXL Base from scratch, I finally got the hang of it. Honestly, the core is just six parts: input text, load model, encode text, run UNET through the sampler, VAE decode, and display the image.
Following along and connecting everything step by step really helped me understand what each slot does. The three outputs on the right side of Load Checkpoint correspond to the three parts of the model. The purple MODEL one goes to KSampler, the two orange ports for positive and negative prompts each need a CLIPTextEncode, and then clip connects back to the loader. The pink latent_image port goes to EmptyLatentImage—for SDXL, remember to set it to 1024x1024. Finally, LATENT goes to VAEDecode, then to the preview.
The coolest part? ComfyUI lets you just drag someone else’s generated image right into the canvas and restore the entire parameter setup. Makes learning other people’s workflows super easy.
Can’t wait for part two on SDXL conditioning.