🐰 Upscaler - McBoaty [1/3] - v5 /u

McBoaty is the upscaling set from the MaraScott pack that's designed around one idea: for print, you want more pixels without the image becoming a different image. The v5 generation is the stable, proven version of that idea, split into three nodes. This one - Upscaler [1/3] - is the front door: it upscales your input and slices it into tiles, then hands everything downstream the per-tile context they need. It does not refine anything itself; that's the Refiner's job in step 3.

What it does

Feed it your image plus the usual model, clip, vae, and positive/negative conditioning (in v5 these are real CONDITIONING inputs - wire in your CLIP Text Encode, unlike v6 which takes text strings). Pick an upscale_model from your models/upscale_models/, set the tile_size (512 default), and the node:

1. Upscales the image with your chosen ESRGAN-style model.
2. Slices it into overlapping tiles (feather_mask controls the blend overlap, default 64px).
3. Stamps each tile with your prompts and sampling parameters for the Refiner to use later.
4. If tile_prompting_active is on, it'll caption each tile with a vision model (Florence-2) and try to give each tile its own prompt - the README calls this experimental, and it is.

You get three outputs:

• McBoaty Pipe - into the Refiner.
• McPrompty Pipe - into a TilePrompter if you want per-tile prompt editing before refining.
• info - run info string.

The settings that matter

The README is blunt about the sweet spots, and they're worth internalizing:

• denoise around 0.27–0.35 - this is the "minimal modification" zone. Higher denoise = more freedom for the model, more seam risk.
• color_match_method - none, mkl, hm, reinhard, mvgd, and hybrid combos. This pulls the refined output's colors back toward the input after refining, which is half of why McBoaty's output doesn't drift into "AI recolored this." The author's own benchmarks brag about this feature.
• Multiple x2 passes beat one x4 - the README says it explicitly. Two 2x upscalers give better print results than a single 4x model.
• vae_encode / tile_size_vae - tiled VAE encode for big images. Keep it on.

One quirk worth knowing: the v5 Upscaler's positive/negative conditioning are required but not used at this stage - the README says they'll be replaced by text prompts in the future. They ride along in the pipe for the Refiner to actually sample with.

Install & ecosystem

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

Restart, or ComfyUI Manager → "MaraScott". You need an upscale model file (4x-UltraSharp, Remacri, etc.) and ideally a canny ControlNet for the Refiner's anchoring. The pack vendors its own Florence-2/Ollama plumbing, so the tile-captioning needs no separate install - just a first-run model download.

Where does this sit in the bigger picture? McBoaty is squarely in the "tiled diffusion upscaling" family - same territory as Ultimate SD Upscale and Tiled Diffusion + ControlNet Tile, but built as one opinionated pipeline with a per-tile editing pass that most tools don't have. It's not the fastest path (the author's own benchmark puts it slower than Ultimate SD Upscale) and not the highest-detail path (SeedVR2 owns that now), but for controlled upscaling that stays faithful to the source, it's a genuinely different, print-oriented take. Start here if you want the full McBoaty control; the Large Refiner collapses steps 1+3 into one node if you want less ceremony.