NAI Enhance 增强

Every generation platform has the "make it better" button, and NAI Enhance 增强 (Enhance) is NovelAI's, given enough dials to be interesting. It takes an image, re-processes it through the model as an img2img run with enhance-specific logic, and returns a sharpened, detail-increased version - optionally bigger. It's the one-button "give it a final pass" that the Enhance feature does on NovelAI's own site, rebuilt as a ComfyUI node you can drop anywhere in a graph.

The inputs are the full house - image, nai_pipe, positive/negative prompts, both settings cards, seed, max_anlas - plus the enhance-specific controls:

• mode - 简化强度 (simplified intensity) or 高级参数 (advanced). Simplified is the sane default: one dial instead of two.
• magnitude - 1–5, default 2. In simplified mode this maps to a hidden (strength, noise) pair - the pack just looks it up. 2 is a light polish, 5 is aggressive.
• strength / noise - only used in advanced mode, so you can set them directly (strength default 0.4). If you've tweaked strength before, you know what you're doing here.
• enhance_scale - the interesting one: 1×, 1.5×, 2×, Max✨（仅 V5）, or 普通缩放. The fixed scales re-render at that multiple; Max✨ is V5's flagship supercharge and only works on V5 - the node checks and errors before billing if you're on an older model. 普通缩放 uses the upscale_percent slider (100–200%) instead.
• upscale_percent - only read in 普通缩放 mode.

It also caps output: any Enhance scale that would exceed NovelAI's 3,145,728-pixel ceiling (the same cap as their official 2× upscale) refuses to run. That's the pack watching out for you at the exact moment a big enhance would get expensive.

The honest framing

Enhance is a detail pass, not a repair shop. If the source is already sharp, an Enhance pass changes the image without improving it - the same "you paid render time for nothing" lesson from the detailing docs applies. The right workflow is: get the image right first, then Enhance as the finishing step. And because it's a paid img2img-style call, the 费用与统计 output matters - it tells you what the pass cost and which seed ran. Wire it to a display node and keep an eye on it; repeated Enhance passes on a big image is a fast way to watch Anlas disappear.

It also rides the same model-capability rails as everything else: the pipe's model decides what's possible (Max✨ = V5 only, for instance), and the node refuses rather than silently downgrading. No Vibe/precise-reference attachments on V5, same as the rest of the pack.

Install

Standard: search ComfyUI-NovelAI-GENYTOOLS in ComfyUI Manager or git clone https://github.com/XTOGENY/ComfyUI-NovelAI-GENYTOOLS.git into custom_nodes, pip install -r requirements.txt, token in ComfyUI settings. Start with simplified mode and magnitude 2, see what the stats line says it cost, then push the dials from evidence rather than optimism.