Following FollowFox’s SDXL ComfyUI series up to the advanced tutorial, here are a few counterintuitive things I noted.
The most confusing part is img2img. You’d think you just swap the empty latent for the image’s latent and set a denoise value, but the advanced KSampler doesn’t even have a denoise input box. Later I realized—skipping the first few steps is equivalent to adding denoise. Total steps determine how much noise you add to the starting image and then remove; skipping the first few steps reduces that amount. So 60 steps skipping the first 30 equals 60 steps with 0.5 denoise.
Pairing it with a refiner makes it even more confusing, because the start/end steps already have their own calculation. Example: txt2img is 25 steps (20 base + 5 refiner). If you want 0.75 denoise for img2img, you’d set 40 steps (25/0.75) and skip the first 15 steps. All these numbers are calculated and linked using SimpleMath nodes.
For LoRA, using Comfyroll’s CR Load LoRA is way better than the standard loader—you can toggle it on/off without deleting the strength value every time. ControlNet is just the canny preprocessor connected to Apply ControlNet, nothing tricky there.