❌ Large Refiner - McBoaty v2 /u

This is "Large Refiner - McBoaty v2," the middle step in the single-node era of McBoaty. If v1 was the proof that a Treeshark12 workflow could be folded into one node, v2 is where it got respectable: the same combined upscale-and-refine pipeline, but with proper sampler choices, sigmas handling, and a tiles output so you can actually see what you're about to refine. It's an old class name now - today's pack registers the current single-node LargeRefiner as v3 and the split three-piece set as v4/v5 - but if an older workflow references this class, it still loads and still works.

How it works

One node does what a manual graph would need five or six nodes to do. It takes your image, upscales it with the model in upscale_model (from models/upscale_models), slices it into tile_size (512) tiles with feather_mask (64) overlap, runs a KSampler pass over each tile with the conditioning you feed in, and stitches everything back together.

Compared to v1, the changes that matter:

• tiles output added - a preview of the tiles before refining, which is what makes the workflow debuggable.
• sampler_name grew to the full 27-choice list and sigmas_type (BasicScheduler / SDTurboScheduler / AlignYourStepsScheduler) plus ays_model_type (SD1/SDXL/SVD) arrived, so you can run turbo/refiner-friendly schedules instead of being stuck with basic.
• The default feather_mask dropped from 350 to 64 - a saner overlap for SDXL-era tiles.

The inputs you'll touch

• output_size (boolean, default true) - unlike the later v3, this is a switch, not a coefficient: true outputs at the upscale model's native size. There's no 1.5x knob here; that arrives in v3.
• vae_encode (true) and tile_size - tiled VAE encoding, so a big upscale doesn't OOM.
• denoise (default 0.35 here, vs 0.27 in v3) - the sweet spot stays in the 0.27–0.35 band. The community's original write-up on this family is blunt about what happens above that: "be prepared of seams and hallucinations."
• positive / negative conditioning - genuinely used by this node, unlike the v4+ Upscaler where they're mostly passthrough. Wire them from your CLIP Text Encodes.

Outputs

image (the refined upscale), tiles (the tile map), original_resized (the pre-refine reference for before/after comparison), and info (a STRING of diagnostics). Wire image to SaveImage and you're done.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/davask/ComfyUI-MarasIT-Nodes

Restart ComfyUI. Or Manager → search "MaraScott" → Install → restart. No model files required; the upscale_model dropdown just needs something like 4x-UltraSharp in models/upscale_models.

One honest note

v2 has no ControlNet inputs and no color-match - those land in v3. If you're building a print upscale today, start from the v3 LargeRefiner (still in the current pack as MaraScottMcBoatyUpscalerRefinerNode_v3) or the v4 trio; reach for this only when you're maintaining an existing workflow that already uses it. It's stable, it's fast, and it's been superseded - both of those can be true at once.