ComfyUI Node

NAI Vibe Transfer

An IP-Adapter you rent by the Anlas

By raspie10032·Created 2 years ago·Updated about 15 hours ago· 10
NAI Vibe Transfer
  • image
  • vibes
  • vibes
◄strength0.60►
◄information_extracted1.00►
◄enabledtrue►

If you've used IP-Adapter, you already understand the idea here (identity-preservation.md): hand the model a reference image and it pulls style, palette and composition cues out of it into your new generation, without a LoRA and without a ControlNet. The difference on NovelAI is that the extraction happens on their servers, half the models charge you for it, and the whole thing is a bundle - not an image - that you plug into the generator.

NovelAI's community is small and stubborn rather than growing; it sits around ten to twenty reddit threads a month, mostly the same people. That's the audience for this node. It's not for everyone, but if you're in it, this is the only sane way to drive Vibe Transfer from a graph.

One image in, a vibe bundle out

The node takes a single image and returns vibes, typed RS_NAI_VIBES. That socket goes into the vibes input on the pack's generator, img2img, inpaint, or Detailer nodes. It is not an IMAGE - you can't preview it, you can't save it, and nothing outside this pack knows what to do with it.

The two knobs:

  • strength (default 0.6) - how hard the reference pushes. This is the one you'll actually tune.
  • information_extracted (default 1.0) - how much is distilled out of the image into the encoding. The tooltip's warning matters more than the number: on V4/V4.5, changing this value requires a new paid encoding.

Plus two optional inputs: vibes (chain another node's output in here to stack references - up to 16) and enabled (off drops that node's image from the chain without you having to rip out the wires). Setting strength to 0 also omits the image, so both switches do the same job.

The generator has a related toggle worth knowing about: normalize_vibe_strengths, on by default, scales your strengths down when their absolute sum exceeds 1.0. Stack four vibes at 0.6 each and something has to give.

The part with the money in it

V3 is the easy case: the prepared image is sent straight to the API. No encoding step, no separate charge for the encoding.

V4 and V4.5 have to be encoded first, via NovelAI's encode-vibe endpoint. The pack caches the returned encoding in ComfyUI/models/rs_nai_vibes, keyed by a hash of the model, the information_extracted value, and the image bytes. The consequences are worth internalising:

  • Change strength all you like - free, the cache is reused.
  • Change the seed - free.
  • Change the image, the model, or information_extracted - new encoding, new charge (2 Anlas per the author's notes).
  • Each vibe beyond the fourth adds 2 Anlas per generation, again per the author's notes. Check NovelAI's current pricing before you build a six-image vibe stack and queue it fifty times.

The cache is a genuinely nice piece of engineering: it validates a checksum on read, and if a cache file is corrupt it tells you the path while warning that re-encoding may cost Anlas, instead of silently re-buying the encoding. It never stores your token. Just don't expect it to survive a models/ wipe.

What it can't do

This is where people lose an evening. Per the pack's own compatibility table:

  • V5 and the legacy Furry model: no Vibe Transfer at all. Vibe is V3, V4, and V4.5.
  • V4/V4.5 Vibe does not work for inpainting or detailing. If you want a reference in an inpaint pass on V4.5, that's NAI Precise Reference, not this.
  • Vibe Transfer and Precise Reference cannot be combined - NovelAI won't accept both. One feature per generation.

The node fails before it sends anything when you get this wrong, which is the behaviour you want: you get an error explaining the model or the mode, not a silent charge and a weird image.

Install and run

cd ComfyUI/custom_nodes
git clone https://github.com/raspie10032/ComfyUI_RS_NAI_API_Request.git
cd ComfyUI_RS_NAI_API_Request
python -m pip install -r requirements.txt

ComfyUI Manager: search ComfyUI_RS_NAI_API_Request, Install, restart. Then create .env in the pack folder with NAI_ACCESS_TOKEN=your_token_here - Vibe needs the token, because encoding is a server call.

The workflow:

Load Image → NAI Vibe Transfer → (vibes) → NAI Image Generator → Preview Image

All the pack's NovelAI calls share one process-local lock with a floor of two seconds between network attempts, one retry on HTTP 429 honouring Retry-After, and a 10s connect / 300s read timeout. So a graph with several reference nodes queues up in a predictable line rather than hammering the API - the cost is that long chains feel slow, because they are.

Leave limit_opus_free on on the generator until you've decided you want to pay. References are blocked by it before the token is even read, and it will block Vibe requests specifically because encoding and extra vibes cost Anlas even when the base generation would have been free.

CategoryRS_NovelAI_API/References

Inputs (5)

NameTypeDefaultDescription
imageIMAGE—
strengthFLOAT0.600–1—
information_extractedFLOAT1.000–1V4/V4.5 encoding costs Anlas on first use of an image/settings combination; later uses reuse the local cache. Changing this value needs a new encoding.
vibesoptRS_NAI_VIBES—
enabledoptBOOLEANtrue—

Outputs (1)

NameTypeDescription
vibesRS_NAI_VIBES—