qwenMultiAngles☀

Building a multi-view dataset means writing a lot of near-identical prompts that only differ in camera angle. qwenMultiAngles exists to stop you hand-typing those. It assembles a single prompt string in the format <sks> {azimuth} {elevation} {distance} - for example <sks> front view eye-level shot medium shot - from three dropdowns, so every training sample gets a consistent, correctly-phrased angle description.

The <sks> trigger and the "front view / eye-level shot / medium shot" vocabulary put this squarely in subject-consistency territory: you're describing the same subject from many camera positions, which is the classic recipe for teaching a model what a thing looks like from every side. It's a small idea, and it saves exactly the kind of repetitive work that makes you sloppy after the twentieth image.

Inputs

• prefix - the trigger, default <sks>. Change it if your training run uses a different token.
• azimuth - 8 options, full compass: front, front-right quarter, right side, back-right quarter, back view, back-left quarter, left side, front-left quarter.
• elevation - 4 options: low-angle, eye-level, elevated, high-angle.
• distance - 3 options: close-up, medium shot, wide shot.

That's all of them. The dropdowns show English plus a Chinese gloss in parentheses (the author's audience is Chinese-language, and the pack's docs are bilingual) - the Chinese is just a label; the English phrasing is what goes into the prompt.

Output is a single prompt STRING, which you'd typically route into whatever's generating your training images, or into a prompt list for a dataset build.

The honest caveats

• It's a prompt formatter, not a dataset generator. It builds the string; you still need the images, the captions, and the training run.
• One prompt per run. There's no multi-output mode, so generating a full 8×4×3 grid means calling it many times or feeding it into a prompt-list workflow.
• The vocabulary is camera-language, not tag-language. If your model wants Danbooru-style tags rather than natural-language camera terms, this isn't the shape you need. For LLM-encoder models like Z-Image and Qwen-Image, natural language is exactly right.

Install

Part of DashuaiTools:

cd ComfyUI/custom_nodes
git clone https://github.com/Hasasasa/ComfyUI_DashuaiTools

then restart ComfyUI, or install via ComfyUI Manager ("DashuaiTools"). Zero dependencies - it's a string concatenation with extra steps.

For its narrow job it's the right tool: consistent, readable, and it removes one more source of typos from dataset prep.