NAI Camera Control

What it's actually for

You've built a NovelAI prompt with character slots - a girl at 0.3, a girl at 0.7 - so you've solved where everyone stands. You haven't solved where the camera is. That's normally the part you fumble by hand: is it from below or from above, do you need dutch angle, and does adding one tag quietly break the framing words you already had?

NAI Camera Control is a dial box that writes those tags for you. Four axes, smooth blending between them, and out comes a caption string already formatted in NovelAI's weight syntax, appended to your scene prompt and - this is the useful bit - to every character caption too. It talks to your NAI context, so the whole thing stays one pipe into the generator.

It's the same engine as RS Camera Control under a different menu entry (RS_NovelAI_API → Camera); the one schema difference is that this one is NovelAI-syntax-first. Local pipeline? Use the other one.

How it works

No ControlNet, no camera model, no extra inference pass. It's arithmetic on text. Each axis is converted into tag weights by blending between knots with a raised cosine, s(t) = (1 − cos(πt)) / 2, so halfway between "front" and "side" you get two tags at roughly half weight each instead of a hard flip. Azimuth blends from front → viewed from the left side / viewed from the right side → from behind across 0°, 90°, 180°, scaled by 0.85 × cos(elevation)². That cosine is why ±90° elevation makes the horizontal caption vanish - at the pole, yaw stops meaning anything. Elevation goes from above / from below at 60°, overhead view / worm's-eye view at 90°. Distance interpolates across five framing tags. Roll adds dutch angle at 0.6 × |roll| / 45.

Then strength multiplies the lot and the result is emitted as 1.02::from behind:: (NovelAI) or (from behind:1.02) (ComfyUI), or as bare tags in plain mode. Weights under 0.005 get dropped, which is fine - the curves are smooth, the text isn't.

The honest caveat, straight from the author: this adds captions, not a camera. The numbers describe caption influence, not measured image angles. Side captions can turn each person independently rather than orbiting one shared camera.

The inputs that matter

You'll mostly touch four sliders and ignore the rest:

• azimuth (−180…180, default 0) - horizontal view, draggable in the little diagram.
• elevation (−90…90) - positive looks down, negative looks up.
• distance (0…1, default 0.5) - 0 extreme close-up, 0.5 medium shot, 1 wide shot.
• strength (default 1.2) - start 0.8–1.4. Push it high and you get distorted or split images rather than a stronger camera.

Then scope decides whether the view caption gets appended to character captions as well (scene_and_characters, the default) or only to the common scene (scene_only). Leave it on the default. Regional setups can suppress a caption that only lives in the common prompt, which is exactly the failure mode you don't want to debug at 1 a.m.

weighting=plain drops the numbers for models that don't parse weighted prompts; prompt_syntax only matters when nothing is connected to building_pipe. random_mode (off/balanced/full) overrides the four axes with a seeded view; balanced stays inside −90…90 azimuth and −35…35 elevation, full uses everything. random_seed is camera-only - your image seed is untouched. Set its control to randomize if you want a new view per queue.

Outputs: building_pipe goes into the generator's nai_context (or the local sampler's character pipe), prompt is scene + camera as one string for a plain text path, and camera_prompt is just the captions - handy for debugging.

Install

ComfyUI Manager → Custom Nodes Manager → search ComfyUI_RS_NAI_API_Request → Install → restart. Or:

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

Heads up: requirements.txt pulls torch, segment-anything and scipy for the pack's detailer and tagger. The camera node needs none of it, no model weights and no token - it never sends a request, so only the downstream generator spends Anlas.

Where people get burned

The missing-scene error. If the pipe you connect has no common prompt, the node refuses to run and asks for base_prompt rather than quietly replacing your scene with camera tags. That's a feature. Connect your scene text there.

Two cameras, or an existing view tag. Chaining cameras stacks captions, and entering a context that already says from behind gives you two views arguing with each other. Clean out the old view words first.

The model ignores you anyway. Prompt-only framing loses to conflicting text - looking at viewer, eye colour calls, an expression that faces the camera. People have chased a reliable back view for months. If a rear view won't take, simplify the scene and character captions before cranking strength.

NAI gets centers, not masks. Character positions are sent as coordinates; silhouettes only matter for local round trips. Don't expect a mask-based regional edit on the API side.