NAI 核心参数

Think of NAI 核心参数 (Core Settings) as the KSampler widget panel of this pack, ripped out into its own node so everything downstream can share it. In a normal ComfyUI graph the resolution, steps, CFG and sampler live on the sampler itself. Here they live on this one card, and the card is what the text-to-image sampler, img2img sampler, inpaint sampler, Enhance node and every batch job all read from. One card, ten consumers, change it once. That's the value-node pattern from the plumbing layer applied to an API workflow - and it's the reason you're not re-entering 832×1216 on six different nodes.

The inputs that matter

Eight fields, and you'll touch maybe four of them daily:

• resolution - a preset dropdown covering the sane range: 小型竖图 512×768, 普通竖图 832×1216 (the default), 普通横图 1216×832, 普通方图 1024×1024, up to 大型横图 1536×1024, plus 自定义. Pick a preset and it fills width/height for you.
• width / height - 64 to 4096 in 64-pixel steps. These override the preset when you pick 自定义. Every sampler rounds its input to a 64-multiple anyway, so don't fight it with weird numbers.
• steps - 1–50, default 23. NovelAI's own UI defaults around here for V5; the API caps single-request steps well below what you'd run locally, and Opus free quota only applies up to 28.
• guidance - 0–10, default 7.0. The classic CFG dial; on NovelAI's models, lower reads more "artistic" and 7 is a perfectly normal place to sit.
• sampler - 欧拉祖先采样 (Euler Ancestral) is the default, with 欧拉, DPM++ 2M, DPM++ 2S 祖先采样, DPM++ SDE, DPM2, DPM 快速 and DDIM behind it. These are NovelAI's sampler names, matching their site, not a free-for-all.
• noise_schedule - 卡拉斯 (Karras) by default, plus 指数, 多项式指数 and 原生. This is the noise schedule, and on NovelAI it matters more than people expect - it's a real lever for how "clean" or "grainy" the render reads.

The outputs

It emits one NAI_CORE_SETTINGS object - wire it into any sampler, the Enhance node, the batch executor, or the character region editor (which reads width/height from it to draw its canvas). There's also a 尺寸摘要 (size summary) string like 832×1216｜23 步｜Guidance 7 so you can read the active config at a glance, or pipe it into a text display node.

Install and notes

Same as the rest of the pack: search ComfyUI-NovelAI-GENYTOOLS in ComfyUI Manager, or git clone https://github.com/XTOGENY/ComfyUI-NovelAI-GENYTOOLS.git into custom_nodes and pip install -r its requirements. You need your NovelAI Persistent API Token in ComfyUI settings before anything actually generates.

The traps are mostly about consistency rather than the node itself. The img2img sampler has a size_mode that can ignore these dimensions and keep the input image's size instead - if your img2img output doesn't match what you set here, that's why. And the tiled-refine node and batch executor each grab the card and stamp per-tile/per-task dimensions on top of a copy, so don't be alarmed when the stats line shows a different size than your card. The card is the seed configuration, not a hard lock.

One more: steps and guidance are per-request and NovelAI bills partly on resolution and steps, so a careless 1536×1024 @ 50 steps card is how you burn through Anlas on a batch job. Set the card high for the one-off you care about, low for the experiments.