π° UpScaler Refiner - McBoaty /u
The single node that started the whole family
- image
- model
- vae
- positive
- negative
- image
- original_resized
- info
Every node in the McBoaty family traces back to this one: "UpScaler Refiner - McBoaty," the v1 that packed an upscale-plus-per-tile-refine pipeline into a single node back in spring 2024. The story matters because it explains why the node is shaped the way it is. It's not a design that was dreamed up in a vacuum - it's a conversion of Rob "TreeShark" Adams' advanced upscaling workflow (there's a YouTube video where it all started), turned into a ComfyUI node by davask (MaraScott), and announced to the world in a pair of reddit posts by Treeshark himself. The tagline in those threads: "Twice as fast as Ultimate Upscale, with, I feel, better quality." That was the whole pitch, and it's still the honest summary.
How it works
Feed it an image, a model, vae, and positive/negative conditioning. It upscales with whatever's in upscale_model (drop 4x-UltraSharp or similar into models/upscale_models if the dropdown is empty), slices the result into tile_size (512) tiles with a generous feather_mask (350 by default in this version - later nodes shrank it to 64), runs a low-denoise KSampler pass over each tile, and re-stitches. The output_size input here is a boolean switch (default true) meaning "output at the upscale model's size" - the fancier coefficient version didn't exist yet.
Sampling is where v1 is clearly a first draft: sigmas_type and ays_model_type (SD1/SDXL/SVD) exist, but the sampler list is shorter and there's no ControlNet, no color match, no tiles output. denoise defaults to 0.35, which the author's own guidance frames as the top of the safe band - 0.27 to 0.35 is where this pipeline behaves; above that, "be prepared of seams and hallucinations."
Outputs
Three wires: image (the refined upscale), original_resized (the upscaled-but-unrefined reference for comparison), and info (a STRING with run diagnostics). That's it - you can't inspect the tiles or feed them anywhere, which is the biggest thing v2 fixed.
Installing it
It's in the same pack, registered under the old "MarasIT" name - install once and it comes along:
cd ComfyUI/custom_nodes
git clone https://github.com/davask/ComfyUI-MarasIT-Nodes
Restart ComfyUI. Manager route: search "MaraScott" β Install β restart.
Should you use it in 2026?
Honestly? Only if you're loading an old workflow. v1 is the least-featured node in the family: no ControlNet to keep tiles faithful, no color match to fix drift, no tile preview. If you're building fresh, use the v3 LargeRefiner or the v4 Upscaler + TilePrompter + Refiner set - same idea, with the tools that actually make tiled upscaling reliable at print sizes. But there's a real argument for keeping one of the old ones around: they're the smallest, simplest possible demonstration of the whole per-tile-refine concept, and they run on a mid-range card without complaint. Fine for learning, superseded for shipping.
Inputs (18)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | β | |
| output_size | BOOLEAN | true | β |
| upscale_model | COMBO | 0 options: | |
| feather_mask | INT | 3500β16384 | β |
| model | MODEL | β | |
| vae | VAE | β | |
| vae_encode | BOOLEAN | true | β |
| tile_size | INT | 512320β4096 | β |
| seed | INT | 40β18446744073709550000 | β |
| steps | INT | 101β10000 | β |
| cfg | FLOAT | 2.50β100 | β |
| sigmas_type | COMBO | 3 options: BasicScheduler, SDTurboScheduler, AlignYourStepsScheduler | |
| sampler_name | COMBO | 22 options: euler, euler_ancestral, heun, heunpp2, dpm_2, dpm_2_ancestral, +16 | |
| basic_scheduler | COMBO | 6 options: normal, karras, exponential, sgm_uniform, simple, ddim_uniform | |
| ays_model_type | COMBO | 3 options: SD1, SDXL, SVD | |
| positive | CONDITIONING | β | |
| negative | CONDITIONING | β | |
| denoise | FLOAT | 0.350β1 | β |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | β |
| original_resized | IMAGE | β |
| info | STRING | β |