Boyo SuperCool Upscaler (Memory Optimized + Batching)

The honest pitch

This node is a ComfyUI wrapper for Andrew DalPino's SuperCool, a fast single-image super-resolution model, and it lives on the pixel rung of the upscaling ladder - not the detail rung. It takes an image that already has the detail you want and makes it bigger, cleanly, in milliseconds rather than minutes. It does not invent pores, lashes, or fabric weave the way SeedVR2 or SUPIR do, and it will never rewrite a face. That is a feature: a 4x pass here changes nothing but resolution, so it won't "human-to-lizard" anyone on a close-up the way a generative restorer can.

The wrapper's actual selling point is the second half of the display name: Memory Optimized + Batching. It auto-unloads the model, runs under mixed precision, and chops big frame sequences into chunks internally. That makes it a genuinely decent cheap upscale stage in a video workflow - feed it a batch of VHS frames and it hands back a batch of 4x frames without you building a tiling rig. The memory improvements were contributed by Maelstrom2014, and they're the reason the thing is usable at all on an 8GB card.

Fair warning before you get attached: this is an obscure tool. A search of the usual community corners turns up essentially nothing about SuperCool or this pack, so treat it as a utility you adopt on your own terms, not a community darling. It fills a narrow slot and it fills it quietly.

How it works

The node builds the SuperCool architecture on the fly: a deep low-resolution conv encoder with SiLU activations, a sub-pixel (pixel-shuffle) decoder, and a bicubic skip path that adds the upscaled original back as a global residual. It's the same family as ESRGAN - classic learned super-resolution, no diffusion involved.

Two things trip people up, both baked into the code:

• Every model is 4x. All of them. The Small (4x), Medium (4x), and Large (4x) dropdown entries map to files named supercool_2x.safetensors, supercool_4x.safetensors, and supercool_8x.safetensors, but the internal upscale ratio is hardcoded to 4 for all three. The names are just ComfyUI dropdown compatibility, as the README admits. The sizes trade speed for quality - Small is 64 channels and 16 layers, Large is 256 channels and 32 layers - but the scale factor never changes.
• It's not a tiler. For images bigger than a few megapixels, upscaling is a RAM/VRAM problem, and this node doesn't tile. If you need 4K+ from a small card, this isn't your tool; that's Ultimate SD Upscale with ControlNet Tile territory.

The inputs that matter

Three of the five inputs actually need your attention:

• model_type - Small / Medium / Large. Start with Small; it's fast and the quality gap is smaller than the name implies. Step up if you're doing a final render.
• batch_size (default 10) - how many images the node processes per internal chunk before cleaning up. The README's rough guide: RTX 3060 (8GB) → 5–8, 4070/4080 → 10–15, 4090 → 20–30+. Drop it if you're also holding a video model in VRAM.
• unload_after_run (default on) - unloads the model and empties the CUDA cache after the pass so downstream nodes get their VRAM back. Leave it on unless you're chaining multiple upscales and want the cache hit.

The other two - image in, a single IMAGE tensor out - wire straight into whatever you're feeding: a Load Image, the output of a VAE decode, or a Video Helper Suite frame batch. There's nothing else in the optional-inputs cupboard; this node is deliberately skinny.

Install

Two ways, both from the README:

# ComfyUI Manager: search "BoyoSupercoolWrapper" and install
# or, manually:
cd ComfyUI/custom_nodes
git clone https://github.com/DragonDiffusionbyBoyo/BoyoSupercoolWrapper

Restart ComfyUI. There are no extra pip dependencies - the wrapper runs on ComfyUI's own torch, safetensors, and huggingface_hub. The real install step is the models:

1. Grab the checkpoints from HuggingFace - andrewdalpino/SuperCool-4x-Small, -Medium, and -Large (or just one, honestly).
2. Rename them supercool_2x.safetensors, supercool_4x.safetensors, supercool_8x.safetensors.
3. Drop them in ComfyUI/models/upscale_models/SuperCool/ (create the folder if needed).

Common issues

• "Model file not found" - the most common failure by far, and it's a file problem, not a node problem. The node looks in models/upscale_models/SuperCool/ for exactly the three names above; a model named supercool_4x.safetensors in the upscale_models root will not be found. Check the folder and the rename.
• Out of memory on batches - that's the batch_size knob, not a bug. Drop it until the run fits. The whole point of the internal batching is that you tune one number instead of restructuring the workflow.
• Expecting an 8x upscale from the "Large" preset - see above. Every preset is 4x. If you need to go further, chain the node twice (4x × 4x = 16x), and don't expect the second pass to add anything the first one couldn't.

One more thing worth knowing: the author cheerfully labels themselves a "vibe coder" in the README, so this pack is maintained with enthusiasm rather than formality. It's simple enough that it mostly doesn't matter - but pin your ComfyUI version if you depend on it.