One click, whole workflow
Recipes are here. Package a whole workflow and run it in one click
Explore recipesOne click, whole workflow
Recipes are here. Package a whole workflow and run it in one click
Explore recipesbyAlibaba & dx8152
Rotate camera angles, shift elevations, and synthesize 3D perspective views from a single still photograph
Transform single 2D images into multi-angle perspectives using precise horizontal rotation, vertical pitch, and forward camera movements.
Built on a 20B dual-encoder architecture trained on 3D Gaussian Splatting, delivering consistent parallax, occlusion, and lighting recalculation.
Explore how camera shifts across products, portraits, and cinematic compositions maintain subject identity while rendering realistic 3D depth.
From e-commerce multi-view galleries to cinematic storyboard adjustments, creators use spatial camera synthesis to bypass full 3D rendering pipelines.
Yes, Qwen Edit Angles is powered by the Qwen-Edit-Multiple-angles LoRA adapter built on top of the 20-billion parameter Qwen Image Edit foundation model. The adapter was trained on over 3,000 synthetic image pairs generated from 3D Gaussian Splatting to map 2D images into consistent 3D camera coordinate spaces.
You can adjust horizontal rotation from -90 to +90 degrees, vertical pitch from -1 (bird's-eye view looking down) to +1 (worm's-eye view looking up), forward dolly zoom from 0 to 10, and toggle wide-angle lens distortion. The model recalculates occlusion, highlights, and casting shadows based on the new virtual camera position.
Use 6 to 8 inference steps with an effect strength between 0.9 and 1.1 for standard rotations while preserving subject identity. Lower the effect strength to 0.7–0.8 if you need to prioritize fine background details, or raise it to 1.3–1.5 if the camera angle requires a more aggressive transformation.
Avoid source images with fine branding typography, intricate wire-like meshes, or noisy low-contrast subjects, as thin details can distort during perspective shifts. Additionally, full 180-degree rotations of highly asymmetric subjects will cause the model to hallucinate or mirror occluded details.
The foundation model was developed by Alibaba, while the specialized multi-angle LoRA adapter was created by dx8152. It pairs a Qwen2.5-VL semantic encoder with a detail-preserving VAE encoder to execute pseudo-3D view synthesis directly on 2D images.
Rotate camera angles, shift elevations, and synthesize 3D perspective views from a single still photograph