Mikey Sampler Base Only

Almost every SDXL checkpoint on CivitAI these days is base-only - no matching refiner model, because the community decided years ago that the refiner stage wasn't earning its keep. That leaves Mikey Sampler's full base→refiner→upscale→base pipeline with nowhere to plug in half its inputs. This is the version built for that reality: same four-stage idea, minus the refiner entirely.

How it works

Instead of base→refiner→upscale→base, this one does base→base→upscale→base - the same model runs twice before the upscale, once fully and once with a step skipped in the middle. That skipped step isn't wasted; it gets added back at the end instead, which the author found adds noticeably more fine detail to the final image than just running a straight number of steps up front. If you've got a plain background or a simple, low-detail region and start seeing faint spotty artifacts from that skip-step trick, smooth_step is the escape hatch - usually one extra smoothing step cleans it up, and for a genuinely busy, detail-heavy image you can push it to -1 for even more correction.

The inputs and outputs that matter

• base_model, samples, positive_cond_base, negative_cond_base, vae - the base essentials, no refiner conditioning needed anywhere.
• model_name - the neural upscale model for the middle stage, same as the full Mikey Sampler: a real ESRGAN-style checkpoint pulled from your upscale_models folder, not a resize algorithm.
• seed - drives the sampling.
• upscale_by (default 1) and hires_strength (default 1, range 0–2) - same knobs as the full sampler: how much bigger the image gets, and how hard the final base pass works on it. 1 is the tested sweet spot for avoiding scrambled detail.
• smooth_step (default 0, range -1 to 100) - extra steps added at the end to clean up the skip-step artifact. Start at 1 if you see spotty backgrounds; -1 for busy images that want more correction than a single positive step gives.

Output is a single LATENT, same as the full sampler - decode with your VAE.

How to install it

Comes bundled with the rest of Mikey Nodes. ComfyUI Manager: search "Mikey Nodes," install. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/bash-j/mikey_nodes

then restart. As with the full sampler, you need an actual upscale model in ComfyUI/models/upscale_models - OpenModelDB is where to get one. The author's pick for a balanced, not-over-sharpened result is lollipop; go with 4x-UltraSharp or NMKD Superscale if you want more visible sharpening.

Common issues & troubleshooting

Spotty artifacts on plain backgrounds. This is the one failure mode the README calls out by name, and it's specific to this node's skip-then-add-back design. smooth_step at 1 fixes it in most cases; if it doesn't, push toward -1 rather than stacking multiple small positive values.

Empty upscale model dropdown. Same cause as every node in this pack that has one - model_name reads from ComfyUI/models/upscale_models, and an empty folder means an empty list. Download something into it.

Scrambled faces at the final size. Lower hires_strength first. It's the same control as the full Mikey Sampler and the same failure mode: too much detail added at once on a region that didn't need much.

Reaching for this on a non-SDXL model. The whole pipeline - resolution assumptions, the base/base/upscale/base structure - is built around SDXL. It won't behave sensibly on Flux, Z-Image, or SD 1.5 checkpoints.