💪TB | Tiny Upscaler (advanced)

The plain Tiny Upscaler keeps all its decisions in genparams. 💪TB | Tiny Upscaler (advanced) throws that away and hands you the steering wheel: you supply the sampler, the sigmas, the conditioning, and the CFG yourself. Same upscaling engine underneath - bilinear scale-up, tiled latent encode, refiner denoise with cross-tiling to avoid grid artifacts - but now with every knob on the dashboard.

If the standard node is "press the upscale button," this one is "build the upscale pipeline." That's a feature if you want to experiment with how the refiner pass behaves, and a liability if you just wanted a bigger image, because you now have to wire up roughly half a sampler yourself.

What you must provide

• image - the thing being upscaled.
• model and vae - the refiner-capable model and VAE. If either is missing, the node logs a warning and passes the image through untouched. Same prototype1 requirement as the basic upscaler.
• positive and negative - actual CONDITIONING tensors, not text. Encode your prompts upstream (with the pack's CLIP or a standard CLIPTextEncode) and feed them in.
• sampler - a SAMPLER, e.g. from ComfyUI's KSamplerSelect. The tooltip literally points you there.
• sigmas - the noise schedule for each denoising step, from any node that outputs SIGMAS (like BasicScheduler).
• scale_by - 1.5× to 6×, default 3.
• enabled - master switch. Off = straight passthrough, no compute.
• cfg - CFG scale, 0 to 15, default 4. Higher sticks closer to the conditioning; lower is looser and more natural.
• extra_noise - 0 to 1.5, default 0.6. This is the noise injected into the upscaled latent before the refine pass, per the README's --upscale-noise parameter. More noise = more detail invention; zero = a cleaner but softer result.
• seed - noise seed for the added noise, with control_after_generate support.

Output: one upscaled image.

How to think about the two noise knobs

This is where beginners get lost, and the tooltips don't help because both cfg and extra_noise carry a copy-pasted CFG description. Read them as different jobs: cfg governs prompt adherence during the refine pass, while extra_noise governs how much fresh detail the upscaler is allowed to invent. For a subtle "was always this resolution" look, keep extra_noise modest; if you want the upscaler to really sharpen and re-detail, that's the knob to raise.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/martin-rizzo/ComfyUI-TinyBreaker

or via ComfyUI Manager (search "tinybreaker"), restart. No pip dependencies. Model files: tinybreaker_prototype1.safetensors (3 GB) in models/checkpoints, t5xxl_fp8_e4m3fn.safetensors (4.9 GB) in models/clip. If you're not actively experimenting with sampler/sigma combinations, the basic Tiny Upscaler is the one to reach for - advanced is for people who want to know why the upscale looks the way it does.