NovelAI 云端生图

The one model family you can't download

NovelAI's V4.5 and V5 anime models are genuinely good, and they're also the ones you'll never run locally. No checkpoints, no Diffusers repo - NovelAI keeps the weights behind its paid cloud API. Want that specific look? Their servers are the only legal route.

NovelAICloudGenerator is the bridge: one node turns a ComfyUI graph into a client for NovelAI's official endpoint, so you keep your habits - saving, upscaling with ESRGAN, batching - while their V5 does the denoising. No GPU work, no checkpoint, nothing to download.

What it actually is

Under the NovelAI category (display name "NovelAI 云端生图"), it covers text-to-image, img2img, and masked inpainting. Heads-up: the whole UI is Chinese, so dropdowns show 欧拉·a (that's Euler a) and 完整版/策展版 (Full/Curated). Don't panic - the fields map cleanly onto NovelAI's website options.

You need a NovelAI account with a subscription or Anlas credits plus a Persistent API Token (pst-...), fetched in the browser under avatar → Account Settings → Get Persistent API Token. No GPU, no model files, no extra Python packages: the pack depends only on ComfyUI's own numpy/Pillow/torch, doing HTTP via the standard library.

How it works

Every generation is an HTTP POST to api.novelai.net with your token as a Bearer header; the finished PNG comes back over the wire. NovelAI encodes your prompt server-side, so your local CLIP model is irrelevant - which is why the CONDITIONING inputs (positive, negative_cond) do something sneaky. A CONDITIONING tensor only carries encoded numbers, no plaintext, so the node walks the execution graph backwards to the upstream encoder's text widget and reads the literal prompt from it. Wire up a CLIP Text Encode and its text replaces the built-in boxes; if no readable text is found, it errors clearly instead of guessing.

Because the API returns finished PNGs, there's no latent to grab. The single output is IMAGE - wire it into SaveImage, or an image upscaler (ESRGAN / 4x-UltraSharp) first. Those PNGs also carry NovelAI's full generation metadata, seed included, so runs stay reproducible.

Inputs a beginner actually touches

• token and proxy - your key, plus an optional HTTP proxy (a necessity in mainland China; falls back to the NAI_PROXY env var, and can stay blank under TUN mode).
• model - V5 / V4.5 / V4 (Full and Curated), plus V3 anime and V3 furry. New official releases go into custom_model_id, which overrides the dropdown.
• size_tier × aspect - two-step sizing like the website, resolving to officially-supported resolutions (832×1216, 1024×1024, 1536×1024…). Unsupported sizes fail or silently burn credits; this sidesteps both.
• steps, cfg_scale, positive_text / negative_text, seed, n_samples.
• quality_tags (Standard auto-appends NovelAI's quality words - "best quality, very aesthetic" heritage) and uc_preset (official negative presets; Heavy adds nsfw suppression).
• Optional image / mask / img2img_strength flip it into img2img or inpainting; a mask auto-switches to the matching -inpainting model variant (white = redraw area) and source images get resized to the target dimensions.

Cost, and the part people forget

Generation spends real money (Anlas). The node live-estimates as you tweak any parameter and auto-cancels when your balance can't cover it - no charge, no surprise. Opus subscribers get the first image free at ≤28 steps and ≤1024², and since your account runs one image at a time, batches generate serially with the seed bumped per image. The 429 concurrency lock is handled for you with exponential backoff.

One honest note: it's anime/illustration only. NovelAI is the danbooru-tag lineage that handed anime SD its clip-skip-2 convention, so reach for this when you want that aesthetic - not photorealism. Already generating locally with Illustrious or Anima? This is how NovelAI's tuned output joins the same graph.

Install and troubleshooting

cd ComfyUI/custom_nodes
git clone https://github.com/Bigesila-B/ComfyUI-NovelAI-Cloud.git

Restart ComfyUI - or use ComfyUI Manager → search NovelAI Cloud → Install. No model downloads, no heavy deps, MIT license; the bundled workflows/basic.json skips the hand-wiring.

Where people get burned:

• The token lives in the workflow. Never share one containing this node - anyone with it can spend your Anlas.
• 401 - you regenerated the token and killed the old one, or pasted quotes/newlines. Re-copy the whole pst- line.
• 402 - out of credits. Drop to V4.5 Full at 832×1216 and ≤28 steps (free first image on Opus), or top up.
• 403 - NovelAI's content policy rejected the prompt. Rewrite it; don't try to route around the filter.
• 429 / concurrency lock - handled automatically; nothing to do.
• Network errors or timeouts - check the proxy port and that the proxy client is running; try 7890/7897/7899, or enable TUN and leave proxy blank.
• Wallpaper tier has no square option - you get a clear error, not a wasted generation.