Mikey Sampler Base Only Advanced

Mikey Sampler Base Only hides its sampler settings behind two simple dials - upscale_by and hires_strength - and picks steps, CFG, sampler, and scheduler for you. This Advanced version rips that hood off. Everything the plain version decided for you is now a widget: steps, two separate CFG values, sampler name, scheduler, even the resize method between passes. Same base→base→upscale→base idea underneath, just with the training wheels off.

What's actually different, not just "more options"

The one change worth flagging before the field list: this node has no model_name input. The plain Base Only sampler upscales with a real neural model (ESRGAN-style, pulled from your upscale_models folder); this Advanced version instead exposes upscale_model as a plain resize algorithm - nearest-exact, bilinear, area, bicubic, or lanczos (default). That's a genuine trade, not a strict upgrade: you give up the neural upscale's ability to add texture in exchange for full control over the sampling math on either side of it. If you were expecting the same lollipop-or-4x-UltraSharp upscale behavior from the non-Advanced node, it isn't here - plan on the resize doing less visual heavy lifting and the second sampling pass doing more.

The inputs and outputs that matter

• base_model, samples, positive_cond_base, negative_cond_base, vae - the same base essentials as the plain version.
• steps (default 31) and smooth_step (default 0, range -1 to 100) - the main pass step count, plus the same skip-then-add-back detail trick from the plain Base Only node.
• cfg_1 (default 5) and cfg_2 (default 9.5) - two separate CFG values across the pipeline's two sampling stages, rather than one CFG for everything. For SDXL and its fine-tunes, the community's general CFG range is 5–9, which is roughly where both of these sit - cfg_1 on the lower, gentler end and cfg_2 pushing higher for the second pass.
• sampler_name (default dpmpp_3m_sde_gpu) and scheduler (default exponential) - real sampler/scheduler control. That default pairing is squarely in DPM++-family territory, which is the safe, well-tested choice for DDPM-style models like SDXL - this is not a flow-matching architecture, so there's no reason to reach for Euler-family samplers here the way you would on Flux or Z-Image.
• add_noise (enable/disable) and denoise (default 1) - standard advanced-sampler controls for how much the process starts from scratch versus refining existing latent content.
• upscale_by (default 1), upscale_model (resize algorithm, default lanczos), hires_denoise (default 0.4), hires_steps (default 31) - the sizing and second-pass controls. hires_denoise in particular is the classic hires-fix range: 0.3–0.5 is where you add detail without redrawing the composition, and 0.4 sits right in the middle of that.

Output is a single LATENT.

How to install it

Bundled with the rest of Mikey Nodes. ComfyUI Manager: search "Mikey Nodes," install, restart. Manually:

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

No neural upscale model download needed for this specific node, since it resizes with plain interpolation rather than a model file.

Common issues & troubleshooting

Expecting the ESRGAN-style upscale from the plain node. This is the single most likely point of confusion moving from Base Only to Base Only Advanced - there's no model_name widget here, and upscale_model is a resize algorithm, not a checkpoint. If detail quality drops after switching, that's why; compensate with hires_denoise and hires_steps rather than looking for a model dropdown that doesn't exist on this node.

Soft results at the final size. Try lanczos if you're not already on it - it's the sharpest of the five resize options and the node's own default - and make sure hires_steps and hires_denoise are doing real work rather than sitting near zero.

Spotty backgrounds. Same fix as the plain Base Only node: bump smooth_step to 1, or toward -1 for busier images.

Non-SDXL checkpoint plugged in. The default sampler/scheduler pairing and the whole two-pass structure assume a DDPM-style SDXL model. On a flow-matching architecture (Flux, Z-Image, Klein), Karras-family and exponential schedules are known to fail outright rather than just underperform - swap architectures, swap your whole sampling approach, not just the model.