Build a 3D camera rig for AI image generation, subscription-free

What it does This Gradio app lets you upload an image and generate new views by manipulating a 3D camera in Three.js space. Drag colored handles to set azimuth and elevation, tweak distance on a slider, and the backend—defaulting to Qwen-Image-Edit-Plus via Alibaba Cloud—renders the scene from that angle. It supports batch generation across multiple viewpoints and keeps a session history gallery.

The interesting bit The 3D rig is essentially a visual prompt builder: your drag-and-drop angles map to discrete 8×4 azimuth/elevation presets plus a continuous distance value, which are then fed to the image-editing model. The README is admirably direct about its purpose—recreating paid services like MultipleAngles so the capability stays outside a paywall.

Key highlights

• Bilingual Chinese/English UI with a pluggable backend; swap the infer_camera_edit function in core.py to hook up a different model.
• Discrete 8×4 angle grid for azimuth and elevation, with continuous distance ranging from 0.6 to 1.8.
• Three.js widget loads from a CDN by default, so it needs an internet connection unless you replace the URL with a local file.
• Local deployment is possible but requires downloading roughly 56 GB of models and an NVIDIA GPU with 16 GB+ VRAM.
• API mode needs no GPU and works on Alibaba Cloud DashScope’s pay-per-call plan.

Caveats

• The 3D drag handles snap to the nearest discrete angle because the underlying LoRA only covers 8×4 preset combinations; continuous free-form rotation is not available.
• The Three.js dependency is fetched from a CDN, meaning the widget will not render in an offline environment without manual source changes.
• Precise viewpoint control is a Qwen-Image-Edit-Plus-specific capability; the README warns that other backends may fail to reproduce the requested angle accurately.

Verdict Worth exploring if you want controlled viewpoint generation without a subscription. Skip it if you need arbitrary continuous angles or are unwilling to wrangle an Alibaba Cloud API key (or 56 GB of local weights).