NAI 高级参数

NAI 高级参数 (Advanced Settings) is the sibling of the core card, and it's where NovelAI's personality lives. The core card gives you steps, guidance and resolution - the boring, universal stuff. This card is the eleven toggles and dropdowns that are specific to NovelAI's API, the ones with names like "CFG Rescale" and "Decrisper" that you won't find on a local KSampler. Output is a single NAI_ADVANCED_SETTINGS object that every sampler, the Enhance node and the batch executor consume alongside the core card. Treat them as a pair - core for what you understand, advanced for the stuff NovelAI invented.

The ones you'll actually touch

There are twelve inputs here. Ignore two-thirds of them until you have a reason. The ones that earn their place:

• transparent_background - V5 only. Flips the request into RGBA and gives you a transparent background PNG. This is a genuine NovelAI selling point, and it only works on V5 Full/Curated. If you're compositing characters into a scene, this is the toggle.
• add_quality_tags - default on. This appends NovelAI's quality-tag vocabulary (masterpiece, best quality style) to the positive prompt for you. NovelAI's models genuinely trained on those tags, so the default is right; turn it off if you're hand-managing every token.
• quality_preset - 标准 / 轻量 / 关闭. This is the V5-era quality system (Standard, Light, or off). Note it interacts with add_quality_tags, so leave both alone until you know why you're touching them.
• image_format - png or webp. This sets the format flag that rides along in the pipe and, in default mode, tells the NAI 图片保存器 what to write. It doesn't save anything itself.
• uc_preset - the negative prompt presets (自定义, 轻度, V5 轻度, V5 人物重点, V5 兽人重点, 重度, 解剖与文字). Useful when you can't be bothered writing your own negative.
• model_mode - 普通 / 兽人数据集 / 背景数据集 (standard / furry / background dataset). V5 lets you bias the training-data mix.

The ones you'll Google later

• cfg_rescale (0–1) and cfg_delay (0–10) - NovelAI's answer to over-saturated CFG. Rescale dials back the CFG overshoot; delay holds off guidance for the first N steps. Both are mild "fix the plasticky look" dials.
• decrisper - a boolean that softens over-crisp generations. NovelAI added it to fight the oversharpened default; some people swear by it, some hate it.
• smea_mode - 关闭 / 自动 SMEA / SMEA / SMEA DYN. SMEA is NovelAI's sampling trick, and here's the catch: it only works on Anime V3 / Furry V3. On V4/V4.5/V5 the pack will error or ignore it rather than pretend. If you're on V5, leave it off.
• dynamic_thresholding - another V3-era feature for the same reason. Same caveat.
• normalize_vibes - default on. Normalizes Vibe strengths when you stack multiple style references. Leave it on unless you're deliberately weighting vibes unequally.

Install and the one real gotcha

Search ComfyUI-NovelAI-GENYTOOLS in ComfyUI Manager or git clone it into custom_nodes plus pip install -r requirements.txt; you need your Persistent API Token in ComfyUI settings to generate anything.

The gotcha is model-vs-feature matching, and this card is where it bites. The pack does check capability combos before billing - pick SMEA on V5 or transparent background on V4.5 and you'll get a clear error instead of a silently-wrong image. But that also means a workflow saved for one model can hard-fail after you swap the loader to another. If an "incompatible" error appears out of nowhere, the culprit is almost always a setting on this card that the new model doesn't support.