🐰 Large Refiner - McBoaty [1/3] - v5 /u

The "Large Refiner" is the one-node version of the v5 McBoaty pipeline: it merges the Upscaler [1/3] and Refiner [3/3] into a single box, so you can upscale, slice, sample every tile, and stitch the result back together without wiring the three-node chain. If the three-node set is the precision path, this is the "just run it" path - and the README's own workflow images lead with this node, so it's clearly the intended default for most people.

Why it exists

The three-node chain gives you the TilePrompter in the middle, but if you don't need per-tile prompt editing, chaining Upscaler → Refiner by hand is pure ceremony. The Large Refiner does the important part automatically: it runs the first pass over all tiles for you. That matters because the README spells out a gotcha of the split version - if you use the Upscaler standalone and then a Refiner with tiles_to_process set to specific tiles, you have to remember to run "all tiles" once first. The Large Refiner does that first full pass by default, so it's also the natural base for the "refine only the wild tiles" workflow:

LargeRefiner (first pass, all tiles) → TilePrompter (edit a few tiles) → Refiner (second pass, those tiles)

Inputs & outputs

Inputs are the full Upscaler set plus the full Refiner set - that's why the node looks intimidating at first glance: image, model/clip/vae, positive/negative conditioning, upscale_model, output_upscale_method (bicubic default), tile_size, feather_mask, vae_encode/tile_size_vae, color_match_method, the experimental tile_prompting_active + vision/LLM model pickers, then the sampler block (sigmas_type, model_type, sampler_name, basic_scheduler, steps, cfg, denoise) and the canny ControlNet block (control_net_name, thresholds, strength, start_percent, end_percent).

Outputs mirror the Refiner:

• McBoaty Pipe and McPrompty Pipe - for chaining more passes or a TilePrompter.
• image and image (orignal) - result and input (yes, "orignal" is the typo in the schema; the code's spelling, not mine).
• tiles, prompts, info - per-tile outputs for inspection.

The settings to actually touch

Keep the pack's constants in mind: denoise 0.27–0.35 for faithful output, canny ControlNet at strength 0.4 to anchor tiles, tiled VAE on for big images. Two x2 upscale passes beat one x4. If you're chasing the "pushed it too far" workflow, the README's move is to run this node at denoise 0.76 (with control net), then fix individual tiles via TilePrompter + a second Refiner.

Install

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

Restart or use ComfyUI Manager → "MaraScott". You'll need an upscale model in models/upscale_models/ and, ideally, a canny/union ControlNet. The pack vendors its own Florence-2/Ollama dependencies, and the optional tile-captioning downloads Florence-2 on first use.

Honest verdict: for a first McBoaty run, start here rather than the three-node chain. You lose per-tile prompt editing in the same box, but you gain a working upscale in one graph, and you can bolt a TilePrompter + Refiner onto its pipe outputs later if one tile misbehaves. The three-node path is for when you know you'll be editing tiles; the Large Refiner is for when you want a print upscaler that works on the first try.