QWEN Image Edit 2511 Camera Angles Lora

January 12, 2026By Alan Kent · AI agent architect; building Ordinary Animator

I came across a new Lora for Qwen Image Edit 2511 making it easy to view an image from different angles. Great for posing the starting image for a camera shot.

They had a cool approach - they used Gaussian Splatting to generate 3D views to train the Lora on. Why 96 camera angles?

8 compass directions (0, 45, 90, 135, 180, 225, 270, 315 degrees)
4 heights (low angle, level, high angle, top view; -30, 0, +30, +60 degrees)
3 distances (close-up, medium shot, far shot)

I tried it on my own robot model and was pretty impressed.

Input image:

Input image of the robot character used to test the camera angles Lora

I provided the recommended text prompts to get a view from different angles.

Robot character regenerated from a different camera angle by the Qwen Image Edit Lora

Robot character shown from another camera angle generated by the Lora

Robot character viewed from a further camera angle generated by the Lora

Robot character from yet another camera angle generated by the Lora

All up, pretty good considering.

But not good in enough.

In my case, since its a robot with lots of intricate head details, it was not able to predict the side or rear of the head accurately. Here is my character reference sheet for the head details. The side and rear of the heads above do not match.

Character reference sheet showing the robot head details from multiple sides

Another video from Benji (one of my favorite channels for keeping up - he publishes good quality content quickly after the new feature is available).

Separately, someone shared this tweet with me. The video shows spinning and viewing from any angle.

@EHuanglu on X

As far as I can tell, its a UI on Hugging Face for the above model. It provided a UI to let you pick the camera angle, it its actually limited to the three control axis with the same increments - the same 96 angles. The video, rotating around a character, is probably from first/last frame AI video generation, not the multi-angle model itself.

That is a pity for me - I was hoping to get more control over camera angles. I already have a tool to have a 360 background photo and let the user snip from that image (see a previous blog post).

Hugging Face UI for the Qwen multi-angle 3D camera model with angle controls

Next experiments I am considering:

If I start with a 45 degree image (so more of the front and side of the head is visible, will it improve accuracy for the side and rear of head?
Train a Lora on my model for Qwen Image Edit 2511 and see if it helps with the side and rear of the head. With a human and hair, this might not be needed. With my robot character, I do need it (or turn my robot character into a liquid metal terminator!)
For other angles, pick two angles near what I want and render out a first/last frame video as an image sequence and pick a frame as my starting point. So if I want 22.5 degrees, use a video starting at 0 degrees and ending at 45 degrees, then take a frame near the middle of the video.
Consider integrating with my 360 clip tool, so I can specify direction, angle, and distance values once, and it clips from the 360 background image at the same time as performing the Image Edit shot composition.

All up, I can see myself using this multi-camera Lora for different angles for video clip starting images, but I am hopeful that training a Lora on my own model will improve the details of the original character not in the original image.