🐰 Large Refiner - McBoaty - v3 /u

This is the old man of the McBoaty family. The v3 "Large Refiner" is the node that started it all - a single box that takes an image, upscales it, slices it into tiles, refines each tile with a sampler, and hands you the stitched result. It predates the three-node v5 split, the per-tile prompt editor, and the v6 restructure. Understanding its history explains why everything after it looks the way it does.

Where it came from

The changelog tells the story. V1 was literally a workflow by Rob Adams (Treeshark) - an "Advanced UpScaling in ComfyUI" video - and MaraScott's v2 converted that workflow into a node. V3 is where it got good: dynamic tiling and LLM-assisted prompt generation arrived, and the node became a real product rather than a workflow in a trench coat. The README still names Treeshark in the credits, and the Treeshark-era reddit posts ("Upscaled Using the Maras nodes Upscale/Refiner McBoaty," +13, which noted it ran "twice as fast as Ultimate Upscale" at sane denoise) are part of how the pack built its reputation. So this node is, genuinely, the piece of ComfyUI history that McBoaty grew from.

What it does

Same core as the v5 Large Refiner, with the earlier (v3) input layout:

• image, model/clip/vae, positive/negative conditioning, seed.
• upscale_model plus output_upscale_method (bicubic default) for the resampling path.
• output_size_type / output_size - upscale-size vs a custom coefficient of input size.
• The sampler block - steps (10), cfg (2.5), sigmas_type, sampler_name, basic_scheduler, denoise (0.27).
• ays_model_type - note the name; in v3 the model-family picker is named for AlignYourSteps and offers SD1/SDXL/SD3/SVD (no FLUX1 - that's a v5/v6 addition). You still pick it even with a different scheduler, because it sets the sampler's native tile size.
• tile_size (512), feather_mask (64), vae_encode/tile_size_vae - the tiling geometry and tiled VAE.
• color_match_method, the experimental tile_prompting_active + vision_llm_model/llm_model.
• The canny ControlNet block - control_net_name, thresholds, strength (0.4), start_percent/end_percent.

Outputs are image, tiles, original_resized (the input scaled to match - useful for comparison), and info.

Should you use it today?

Honestly? Probably not as a first choice. v5's Large Refiner is a superset: same single-node convenience, plus the FLUX1 model type, the improved per-tile prompt handling, and the pipe outputs (McBoaty Pipe/McPrompty Pipe) that let you chain refiners and TilePrompters. This v3 node returns plain images with no pipes, so you can't iterate on it. It's the version to reach for if you're running an old workflow that references it, or if you want the simplest possible "one box upscale" with no pipe plumbing to learn.

Install

Same pack, same way:

cd ComfyUI/custom_nodes
git clone https://github.com/MaraScott/ComfyUI_MaraScott_Nodes

Restart or ComfyUI Manager → "MaraScott". Needs an upscale model in models/upscale_models/; a canny ControlNet is optional but recommended for the tile anchoring that keeps McBoaty faithful. One v3-specific note: since it's the oldest active node in the set, if you're on the latest ComfyUI and something in the newer versions of the pack feels unstable, this node being the "last updated" one means it's both the most battle-hardened and the least likely to get new fixes. It's a museum piece that still works - which, for a node that started a whole upscaling family, is a pretty good legacy.