Zura V3 · Pick camera angles

Every multi-camera workflow I've seen dies at the same place: the camera decisions live in your head, and the graph can't see them. You generate nine angle stills, then eyeball which one goes where across a two-minute clip, then try to remember. Zura V3 · Pick camera angles is the node that fixes that - it turns a 3×3 grid of candidate angles into a clickable gallery with a per-shot timeline, and it writes a shot plan the render nodes read.

It's the front door of the pack's V3 multicam flow. Not a generator itself; a planner and a filing cabinet.

What the nine angles are

The grid is camera side by framing: Left / Front / Right across, wide / medium / close down. Those are azimuths of −30°, 0° and +30°, with three "distance" planes. The author's own comment in the source is worth knowing before you build expectations: the comment in multicam_v3.py says CrossView's current LoRA does not follow distance reliably, so shot size is a visual proposal rather than a dolly command. The close row gets its shot size from a native crop downstream, not from asking the model to zoom.

Which means: you feed this node nine images. Where they come from is your business - a multi-angle LoRA on an image editor, a depth-warp guide, whatever you already trust. The node's job starts after that.

Inputs you actually set

• source_frames - the trimmed 24 fps source clip, decoded frames. Everything is measured against these.
• frames_per_shot - default 24, one second at 24 fps. Range 12–3600. This sets how many shots the plan contains: ceil(frames / frames_per_shot), capped at 60.
• shot_choices - a comma list of angle numbers, one per shot, 0 meaning the original camera. The default 0,6,0,4 is a four-shot example.
• project_name - default performance-v3. Sanitised to letters, digits, _ and -, then truncated at 48 characters.
• stage - comes from Zura V3 · Choose pass. In pass 1 the node demands all nine candidate_1…candidate_9 images via ComfyUI's lazy-input mechanism, so nine generations actually run. In pass 2 it doesn't want them at all.
• shot_prompts - a JSON list of strings, one per shot. Only meaningful on generated-angle shots; a prompt on an original-camera shot is an error downstream.
• Optional look_prompt / look_enabled - a hash of the look is stored in the manifest, so changing your studio prompt invalidates the saved previews.

Outputs are shot_plan (the JSON - wire it into the H3 render node's shot_plan and into the V4 clean node), contact_sheet (a 960×645 labelled grid, perfect for Preview Image), and stage (pass the stage back out to other V3 nodes rather than daisy-chaining switches).

How it works

The planning pass writes a folder under ComfyUI/output/zura_multicam_v3/<project>/<signature>/ containing source.png, angle_1.png…angle_9.png and a manifest.json. The signature is a 16-hex digest over the frame shape plus sampled pixels from the first, middle and last frames - so it's specific to that exact clip, and re-trimming the source invalidates every saved angle. That's deliberate, and it's the single most common error you'll hit (The source changed after angle planning. Re-run Plan angles.).

The gallery UI is the good part: click a tile, then click a shot in the timeline. Clicking a camera saves your selection, it does not queue a render - the author calls that out in the UI itself, presumably after someone filed a bug.

Install

ComfyUI Manager → search Zura Nodes (publisher zuravfx), or:

cd ComfyUI/custom_nodes
git clone https://github.com/ZURAVFX/ComfyUI_zura_nodes
pip install -r ComfyUI_zura_nodes/requirements.txt

Restart ComfyUI. ffmpeg and ffprobe need to be on PATH for the pack's video handling. It's AGPL-3.0-only and ships no model weights - your H3 / Klein / Wan models and LoRAs are yours to place.

Where people get burned

shot_choices has to have exactly as many entries as the clip has shots. In planning mode the node is forgiving and fills the gaps with 0s; in the render pass it just fails. If you changed frames_per_shot after picking angles, you have more or fewer shots than choices and the plan no longer describes your clip.

And the honest limit from the README: a structurally valid graph is not visual sign-off. Plan a short clip, render it, and look at the joins before you put an hour of footage through it.