NAI 官方放大

Most upscaling in ComfyUI is free, local, and fast - an ESRGAN model or Lanczos gives you pixels in milliseconds. This node is none of those things. NAIOfficialUpscaleV2 sends your image to NovelAI's servers and runs their official upscaler, which costs Anlas and is fixed at 2×. So the first question is why you'd pay for what you can do locally. The answer: this is NovelAI's own model, tuned to their art style, and for that specific look it's the one upscaler that won't fight the original.

What it is

Inputs are an IMAGE and a scale selector - and the selector has exactly one option: "新版固定 2×（按输入像素计费）" (the new fixed 2×, billed by input pixels). NovelAI's upscale API is currently locked to 2×; the 4× option that existed in older versions of the pack has been migrated to the new 2× behavior, and the pack says so explicitly in its stats output if you were on the old workflow. There's nothing to configure, which is the entire point: you feed it an image and it comes back bigger.

Outputs:

• 图像 (IMAGE) - the upscaled image, ready for a saver or further processing.
• 统计 (STRING) - the cost estimate in Anlas and the output dimensions.

How the pricing works

This is the part that surprises people. The cost is tiered by input pixels, not output size, and it maxes out at 4 Anlas:

• ≤ 1,048,576 pixels (about 1024×1024): 1 Anlas
• ≤ 1,747,627 pixels: 2 Anlas
• ≤ 2,446,678 pixels: 3 Anlas
• up to 3,145,728 pixels (the hard input cap): 4 Anlas

Above 3,145,728 input pixels the node errors out rather than charging you - that's NovelAI's limit, enforced client-side before the request. So upscaling a big image costs more, which is the opposite of the "megapixels per dollar" economy you might expect from local upscalers.

When you'd use this over a local upscaler

The KB's upscaling essay splits upscaling into two different jobs: adding pixels (ESRGAN, Lanczos - free, instant, can't invent detail) and adding detail (generative - slow, hallucinates, rewrites faces). NovelAI's official upscale sits in the middle: it's a model upscale, so it will add detail, but it's the detail NovelAI's own aesthetic expects. If you're deep in the NovelAI art pipeline and you want a bigger image that still looks like the original generation, the official upscaler keeps the style consistent in a way RealESRGAN won't.

Where it loses: anything on a budget. For pure resolution, a local ESRGAN model is free and done in a second. The official node is for when style consistency is worth an Anlas or four.

Install and setup

Part of ComfyUI-NovelAI-GENYTOOLS. Manager search, or:

cd ComfyUI/custom_nodes
git clone https://github.com/XTOGENY/ComfyUI-NovelAI-GENYTOOLS.git
pip install -r ComfyUI-NovelAI-GENYTOOLS/requirements.txt

Restart, then set your NovelAI Persistent API Token under NovelAI 2.0: Persistent API Token. No model downloads - the upscaler runs server-side.

Gotchas

• It spends Anlas, and the cost scales with input size. Upscaling a 2048×2048 image costs more than two 1024×1024 images' worth of generation. Budget accordingly.
• The hard input cap is 3,145,728 pixels. If your image is bigger, downscale it first (or tile the work and upscale tiles - which is a whole other part of this pack).
• scale has one choice and it's not a bug: NovelAI's current upscale API is 2× only. Old 4× workflows were migrated to 2× rather than silently faked.
• It always re-runs (no ComfyUI caching), so if you re-execute the graph it will re-bill. Disconnect the node or bypass it when you're just rerunning downstream.